From 6de1a82575454e74a2db2acaa97e2cfb01e64ab1 Mon Sep 17 00:00:00 2001 From: Adam Wilson Date: Mon, 18 Aug 2025 20:27:56 -0600 Subject: [PATCH 1/5] fix llama --- ...s_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.01-20.yml | 4 ++-- ...s_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.21-40.yml | 4 ++-- ...s_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.41-60.yml | 4 ++-- ...s_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.61-80.yml | 4 ++-- ..._prompts_rag_and_cot_meta_llama_3_2_3b_instruct.81-100.yml | 4 ++-- 5 files changed, 10 insertions(+), 10 deletions(-) diff --git a/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.01-20.yml b/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.01-20.yml index a8e3d4470..b16703246 100644 --- a/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.01-20.yml +++ b/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.01-20.yml @@ -1,10 +1,10 @@ -name: '#4 (1-20) | RAG + CoT | microsoft/Phi-3-mini-4k-instruct' +name: '#4 (1-20) | RAG + CoT | meta-llama/Llama-3.2-3B-Instruct' on: workflow_dispatch: jobs: test-prompts-1-20: - uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_microsoft_phi_3_mini4k_instruct.base.yml + uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.base.yml with: batch_offset: 0 range_name: "1-20" diff --git a/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.21-40.yml b/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.21-40.yml index 56efda8e4..3f5e989fc 100644 --- a/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.21-40.yml +++ b/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.21-40.yml @@ -1,10 +1,10 @@ -name: '#4 (21-40) | RAG + CoT | microsoft/Phi-3-mini-4k-instruct' +name: '#4 (21-40) | RAG + CoT | meta-llama/Llama-3.2-3B-Instruct' on: workflow_dispatch: jobs: test-prompts-21-40: - uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_microsoft_phi_3_mini4k_instruct.base.yml + uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.base.yml with: batch_offset: 20 range_name: "21-40" diff --git a/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.41-60.yml b/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.41-60.yml index 474d66bca..1cd91b23d 100644 --- a/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.41-60.yml +++ b/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.41-60.yml @@ -1,10 +1,10 @@ -name: '#4 (41-60) | RAG + CoT | microsoft/Phi-3-mini-4k-instruct' +name: '#4 (41-60) | RAG + CoT | meta-llama/Llama-3.2-3B-Instruct' on: workflow_dispatch: jobs: test-prompts-41-60: - uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_microsoft_phi_3_mini4k_instruct.base.yml + uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.base.yml with: batch_offset: 40 range_name: "41-60" diff --git a/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.61-80.yml b/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.61-80.yml index 7f078d0a7..f54cf2b05 100644 --- a/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.61-80.yml +++ b/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.61-80.yml @@ -1,10 +1,10 @@ -name: '#4 (61-80) | RAG + CoT | microsoft/Phi-3-mini-4k-instruct' +name: '#4 (61-80) | RAG + CoT | meta-llama/Llama-3.2-3B-Instruct' on: workflow_dispatch: jobs: test-prompts-61-80: - uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_microsoft_phi_3_mini4k_instruct.base.yml + uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.base.yml with: batch_offset: 60 range_name: "61-80" diff --git a/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.81-100.yml b/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.81-100.yml index f44cb09ba..4baf18b1f 100644 --- a/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.81-100.yml +++ b/.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.81-100.yml @@ -1,10 +1,10 @@ -name: '#4 (81-100) | RAG + CoT | microsoft/Phi-3-mini-4k-instruct' +name: '#4 (81-100) | RAG + CoT | meta-llama/Llama-3.2-3B-Instruct' on: workflow_dispatch: jobs: test-prompts-81-100: - uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_microsoft_phi_3_mini4k_instruct.base.yml + uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_meta_llama_3_2_3b_instruct.base.yml with: batch_offset: 80 range_name: "81-100" From fad7035b3296b4887f1f7e4b23170e12646e43aa Mon Sep 17 00:00:00 2001 From: Adam Wilson Date: Mon, 18 Aug 2025 20:38:59 -0600 Subject: [PATCH 2/5] batch test workflows for OpenELM --- ...nd_cot_apple_openelm_3b_instruct.01-20.yml | 11 ++++++++ ...nd_cot_apple_openelm_3b_instruct.21-40.yml | 11 ++++++++ ...nd_cot_apple_openelm_3b_instruct.41-60.yml | 11 ++++++++ ...nd_cot_apple_openelm_3b_instruct.61-80.yml | 11 ++++++++ ...d_cot_apple_openelm_3b_instruct.81-100.yml | 11 ++++++++ ...and_cot_apple_openelm_3b_instruct.base.yml | 27 +++++++++++++++++++ 6 files changed, 82 insertions(+) create mode 100644 .github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.01-20.yml create mode 100644 .github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.21-40.yml create mode 100644 .github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.41-60.yml create mode 100644 .github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.61-80.yml create mode 100644 .github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.81-100.yml create mode 100644 .github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.base.yml diff --git a/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.01-20.yml b/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.01-20.yml new file mode 100644 index 000000000..2b6d2c091 --- /dev/null +++ b/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.01-20.yml @@ -0,0 +1,11 @@ +name: '#4 (1-20) | RAG + CoT | apple/OpenELM-3B-Instruct' +on: + workflow_dispatch: + +jobs: + test-prompts-1-20: + uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.base.yml + with: + batch_offset: 0 + range_name: "1-20" + batch_size: 2 \ No newline at end of file diff --git a/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.21-40.yml b/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.21-40.yml new file mode 100644 index 000000000..753d83445 --- /dev/null +++ b/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.21-40.yml @@ -0,0 +1,11 @@ +name: '#4 (21-40) | RAG + CoT | apple/OpenELM-3B-Instruct' +on: + workflow_dispatch: + +jobs: + test-prompts-21-40: + uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.base.yml + with: + batch_offset: 20 + range_name: "21-40" + batch_size: 2 \ No newline at end of file diff --git a/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.41-60.yml b/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.41-60.yml new file mode 100644 index 000000000..9a6a654eb --- /dev/null +++ b/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.41-60.yml @@ -0,0 +1,11 @@ +name: '#4 (41-60) | RAG + CoT | apple/OpenELM-3B-Instruct' +on: + workflow_dispatch: + +jobs: + test-prompts-41-60: + uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.base.yml + with: + batch_offset: 40 + range_name: "41-60" + batch_size: 2 \ No newline at end of file diff --git a/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.61-80.yml b/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.61-80.yml new file mode 100644 index 000000000..c8b51fdf1 --- /dev/null +++ b/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.61-80.yml @@ -0,0 +1,11 @@ +name: '#4 (61-80) | RAG + CoT | apple/OpenELM-3B-Instruct' +on: + workflow_dispatch: + +jobs: + test-prompts-61-80: + uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.base.yml + with: + batch_offset: 60 + range_name: "61-80" + batch_size: 2 \ No newline at end of file diff --git a/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.81-100.yml b/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.81-100.yml new file mode 100644 index 000000000..8d82586a5 --- /dev/null +++ b/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.81-100.yml @@ -0,0 +1,11 @@ +name: '#4 (81-100) | RAG + CoT | apple/OpenELM-3B-Instruct' +on: + workflow_dispatch: + +jobs: + test-prompts-81-100: + uses: ./.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.base.yml + with: + batch_offset: 80 + range_name: "81-100" + batch_size: 2 \ No newline at end of file diff --git a/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.base.yml b/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.base.yml new file mode 100644 index 000000000..f52be072d --- /dev/null +++ b/.github/workflows/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.base.yml @@ -0,0 +1,27 @@ +name: 'Reusable Test #4 | RAG + CoT | apple/OpenELM-3B-Instruct' +on: + workflow_call: + inputs: + batch_offset: + description: 'Starting prompt index offset' + required: true + type: number + range_name: + description: 'Human readable range name (e.g., "1-20")' + required: true + type: string + batch_size: + description: 'Number of prompts per batch' + required: false + type: number + default: 2 + +jobs: + test: + uses: ./.github/workflows/test_04.abstract_base.yml + with: + batch_offset: ${{ inputs.batch_offset }} + range_name: ${{ inputs.range_name }} + batch_size: ${{ inputs.batch_size }} + test_file_path: tests/integration/test_04_malicious_prompts_rag_and_cot_apple_openelm_3b_instruct.py + model_display_name: apple/OpenELM-3B-Instruct \ No newline at end of file From 28a962facd6ab275a1ba59bb72e4abc88502894b Mon Sep 17 00:00:00 2001 From: Adam Wilson Date: Mon, 18 Aug 2025 20:48:43 -0600 Subject: [PATCH 3/5] support casing --- .../services/logging/test_run_logging_service.py | 2 +- tests/create_dirs.sh | 2 +- tests/logs/test_1/apple_OpenELM-3B-Instruct/1-2/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/11-12/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/13-14/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/15-16/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/17-18/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/19-20/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/21-22/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/23-24/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/25-26/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/27-28/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/29-30/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/3-4/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/31-32/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/33-34/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/35-36/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/37-38/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/39-40/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/41-42/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/43-44/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/45-46/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/47-48/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/49-50/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/5-6/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/51-52/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/53-54/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/55-56/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/57-58/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/59-60/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/61-62/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/63-64/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/65-66/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/67-68/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/69-70/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/7-8/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/71-72/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/73-74/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/75-76/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/77-78/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/79-80/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/81-82/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/83-84/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/85-86/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/87-88/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/89-90/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/9-10/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/91-92/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/93-94/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/95-96/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/97-98/_.json | 0 tests/logs/test_1/apple_OpenELM-3B-Instruct/99-100/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json | 0 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json | 0 .../logs/test_1/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/1-2/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/11-12/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/13-14/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/15-16/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/17-18/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/19-20/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/21-22/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/23-24/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/25-26/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/27-28/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/29-30/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/3-4/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/31-32/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/33-34/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/35-36/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/37-38/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/39-40/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/41-42/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/43-44/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/45-46/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/47-48/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/49-50/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/5-6/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/51-52/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/53-54/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/55-56/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/57-58/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/59-60/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/61-62/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/63-64/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/65-66/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/67-68/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/69-70/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/7-8/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/71-72/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/73-74/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/75-76/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/77-78/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/79-80/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/81-82/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/83-84/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/85-86/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/87-88/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/89-90/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/9-10/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/91-92/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/93-94/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/95-96/_.json | 0 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/97-98/_.json | 0 .../logs/test_1/microsoft_Phi-3-mini-4k-instruct/99-100/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/1-2/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/11-12/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/13-14/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/15-16/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/17-18/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/19-20/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/21-22/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/23-24/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/25-26/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/27-28/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/29-30/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/3-4/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/31-32/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/33-34/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/35-36/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/37-38/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/39-40/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/41-42/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/43-44/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/45-46/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/47-48/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/49-50/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/5-6/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/51-52/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/53-54/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/55-56/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/57-58/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/59-60/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/61-62/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/63-64/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/65-66/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/67-68/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/69-70/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/7-8/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/71-72/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/73-74/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/75-76/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/77-78/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/79-80/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/81-82/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/83-84/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/85-86/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/87-88/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/89-90/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/9-10/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/91-92/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/93-94/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/95-96/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/97-98/_.json | 0 tests/logs/test_2/apple_OpenELM-3B-Instruct/99-100/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json | 0 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json | 0 .../logs/test_2/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/1-2/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/11-12/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/13-14/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/15-16/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/17-18/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/19-20/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/21-22/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/23-24/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/25-26/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/27-28/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/29-30/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/3-4/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/31-32/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/33-34/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/35-36/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/37-38/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/39-40/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/41-42/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/43-44/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/45-46/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/47-48/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/49-50/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/5-6/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/51-52/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/53-54/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/55-56/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/57-58/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/59-60/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/61-62/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/63-64/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/65-66/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/67-68/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/69-70/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/7-8/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/71-72/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/73-74/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/75-76/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/77-78/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/79-80/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/81-82/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/83-84/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/85-86/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/87-88/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/89-90/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/9-10/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/91-92/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/93-94/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/95-96/_.json | 0 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/97-98/_.json | 0 .../logs/test_2/microsoft_Phi-3-mini-4k-instruct/99-100/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/1-2/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/11-12/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/13-14/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/15-16/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/17-18/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/19-20/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/21-22/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/23-24/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/25-26/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/27-28/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/29-30/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/3-4/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/31-32/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/33-34/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/35-36/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/37-38/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/39-40/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/41-42/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/43-44/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/45-46/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/47-48/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/49-50/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/5-6/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/51-52/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/53-54/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/55-56/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/57-58/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/59-60/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/61-62/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/63-64/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/65-66/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/67-68/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/69-70/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/7-8/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/71-72/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/73-74/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/75-76/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/77-78/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/79-80/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/81-82/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/83-84/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/85-86/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/87-88/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/89-90/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/9-10/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/91-92/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/93-94/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/95-96/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/97-98/_.json | 0 tests/logs/test_3/apple_OpenELM-3B-Instruct/99-100/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json | 0 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json | 0 .../logs/test_3/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/1-2/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/11-12/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/13-14/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/15-16/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/17-18/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/19-20/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/21-22/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/23-24/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/25-26/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/27-28/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/29-30/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/3-4/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/31-32/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/33-34/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/35-36/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/37-38/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/39-40/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/41-42/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/43-44/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/45-46/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/47-48/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/49-50/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/5-6/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/51-52/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/53-54/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/55-56/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/57-58/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/59-60/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/61-62/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/63-64/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/65-66/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/67-68/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/69-70/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/7-8/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/71-72/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/73-74/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/75-76/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/77-78/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/79-80/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/81-82/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/83-84/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/85-86/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/87-88/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/89-90/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/9-10/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/91-92/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/93-94/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/95-96/_.json | 0 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/97-98/_.json | 0 .../logs/test_3/microsoft_Phi-3-mini-4k-instruct/99-100/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/1-2/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/11-12/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/13-14/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/15-16/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/17-18/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/19-20/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/21-22/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/23-24/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/25-26/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/27-28/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/29-30/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/3-4/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/31-32/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/33-34/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/35-36/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/37-38/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/39-40/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/41-42/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/43-44/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/45-46/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/47-48/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/49-50/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/5-6/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/51-52/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/53-54/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/55-56/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/57-58/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/59-60/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/61-62/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/63-64/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/65-66/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/67-68/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/69-70/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/7-8/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/71-72/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/73-74/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/75-76/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/77-78/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/79-80/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/81-82/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/83-84/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/85-86/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/87-88/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/89-90/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/9-10/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/91-92/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/93-94/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/95-96/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/97-98/_.json | 0 tests/logs/test_4/apple_OpenELM-3B-Instruct/99-100/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json | 0 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json | 0 .../logs/test_4/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/1-2/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/11-12/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/13-14/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/15-16/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/17-18/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/19-20/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/21-22/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/23-24/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/25-26/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/27-28/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/29-30/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/3-4/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/31-32/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/33-34/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/35-36/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/37-38/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/39-40/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/41-42/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/43-44/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/45-46/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/47-48/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/49-50/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/5-6/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/51-52/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/53-54/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/55-56/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/57-58/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/59-60/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/61-62/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/63-64/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/65-66/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/67-68/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/69-70/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/7-8/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/71-72/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/73-74/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/75-76/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/77-78/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/79-80/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/81-82/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/83-84/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/85-86/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/87-88/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/89-90/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/9-10/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/91-92/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/93-94/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/95-96/_.json | 0 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/97-98/_.json | 0 .../logs/test_4/microsoft_Phi-3-mini-4k-instruct/99-100/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/1-2/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/11-12/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/13-14/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/15-16/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/17-18/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/19-20/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/21-22/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/23-24/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/25-26/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/27-28/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/29-30/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/3-4/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/31-32/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/33-34/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/35-36/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/37-38/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/39-40/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/41-42/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/43-44/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/45-46/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/47-48/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/49-50/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/5-6/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/51-52/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/53-54/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/55-56/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/57-58/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/59-60/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/61-62/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/63-64/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/65-66/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/67-68/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/69-70/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/7-8/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/71-72/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/73-74/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/75-76/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/77-78/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/79-80/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/81-82/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/83-84/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/85-86/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/87-88/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/89-90/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/9-10/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/91-92/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/93-94/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/95-96/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/97-98/_.json | 0 tests/logs/test_5/apple_OpenELM-3B-Instruct/99-100/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json | 0 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json | 0 .../logs/test_5/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/1-2/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/11-12/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/13-14/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/15-16/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/17-18/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/19-20/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/21-22/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/23-24/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/25-26/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/27-28/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/29-30/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/3-4/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/31-32/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/33-34/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/35-36/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/37-38/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/39-40/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/41-42/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/43-44/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/45-46/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/47-48/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/49-50/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/5-6/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/51-52/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/53-54/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/55-56/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/57-58/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/59-60/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/61-62/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/63-64/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/65-66/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/67-68/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/69-70/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/7-8/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/71-72/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/73-74/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/75-76/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/77-78/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/79-80/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/81-82/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/83-84/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/85-86/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/87-88/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/89-90/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/9-10/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/91-92/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/93-94/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/95-96/_.json | 0 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/97-98/_.json | 0 .../logs/test_5/microsoft_Phi-3-mini-4k-instruct/99-100/_.json | 0 752 files changed, 2 insertions(+), 2 deletions(-) create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/1-2/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/11-12/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/13-14/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/15-16/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/17-18/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/19-20/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/21-22/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/23-24/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/25-26/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/27-28/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/29-30/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/3-4/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/31-32/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/33-34/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/35-36/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/37-38/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/39-40/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/41-42/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/43-44/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/45-46/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/47-48/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/49-50/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/5-6/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/51-52/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/53-54/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/55-56/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/57-58/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/59-60/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/61-62/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/63-64/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/65-66/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/67-68/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/69-70/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/7-8/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/71-72/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/73-74/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/75-76/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/77-78/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/79-80/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/81-82/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/83-84/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/85-86/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/87-88/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/89-90/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/9-10/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/91-92/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/93-94/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/95-96/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/97-98/_.json create mode 100644 tests/logs/test_1/apple_OpenELM-3B-Instruct/99-100/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json create mode 100644 tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/1-2/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/11-12/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/13-14/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/15-16/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/17-18/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/19-20/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/21-22/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/23-24/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/25-26/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/27-28/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/29-30/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/3-4/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/31-32/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/33-34/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/35-36/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/37-38/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/39-40/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/41-42/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/43-44/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/45-46/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/47-48/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/49-50/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/5-6/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/51-52/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/53-54/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/55-56/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/57-58/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/59-60/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/61-62/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/63-64/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/65-66/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/67-68/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/69-70/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/7-8/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/71-72/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/73-74/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/75-76/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/77-78/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/79-80/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/81-82/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/83-84/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/85-86/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/87-88/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/89-90/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/9-10/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/91-92/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/93-94/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/95-96/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/97-98/_.json create mode 100644 tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/99-100/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/1-2/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/11-12/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/13-14/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/15-16/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/17-18/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/19-20/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/21-22/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/23-24/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/25-26/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/27-28/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/29-30/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/3-4/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/31-32/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/33-34/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/35-36/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/37-38/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/39-40/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/41-42/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/43-44/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/45-46/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/47-48/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/49-50/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/5-6/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/51-52/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/53-54/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/55-56/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/57-58/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/59-60/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/61-62/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/63-64/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/65-66/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/67-68/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/69-70/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/7-8/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/71-72/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/73-74/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/75-76/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/77-78/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/79-80/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/81-82/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/83-84/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/85-86/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/87-88/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/89-90/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/9-10/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/91-92/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/93-94/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/95-96/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/97-98/_.json create mode 100644 tests/logs/test_2/apple_OpenELM-3B-Instruct/99-100/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json create mode 100644 tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/1-2/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/11-12/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/13-14/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/15-16/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/17-18/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/19-20/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/21-22/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/23-24/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/25-26/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/27-28/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/29-30/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/3-4/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/31-32/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/33-34/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/35-36/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/37-38/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/39-40/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/41-42/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/43-44/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/45-46/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/47-48/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/49-50/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/5-6/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/51-52/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/53-54/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/55-56/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/57-58/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/59-60/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/61-62/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/63-64/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/65-66/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/67-68/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/69-70/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/7-8/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/71-72/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/73-74/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/75-76/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/77-78/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/79-80/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/81-82/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/83-84/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/85-86/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/87-88/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/89-90/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/9-10/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/91-92/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/93-94/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/95-96/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/97-98/_.json create mode 100644 tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/99-100/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/1-2/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/11-12/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/13-14/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/15-16/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/17-18/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/19-20/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/21-22/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/23-24/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/25-26/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/27-28/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/29-30/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/3-4/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/31-32/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/33-34/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/35-36/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/37-38/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/39-40/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/41-42/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/43-44/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/45-46/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/47-48/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/49-50/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/5-6/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/51-52/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/53-54/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/55-56/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/57-58/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/59-60/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/61-62/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/63-64/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/65-66/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/67-68/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/69-70/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/7-8/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/71-72/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/73-74/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/75-76/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/77-78/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/79-80/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/81-82/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/83-84/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/85-86/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/87-88/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/89-90/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/9-10/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/91-92/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/93-94/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/95-96/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/97-98/_.json create mode 100644 tests/logs/test_3/apple_OpenELM-3B-Instruct/99-100/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json create mode 100644 tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/1-2/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/11-12/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/13-14/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/15-16/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/17-18/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/19-20/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/21-22/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/23-24/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/25-26/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/27-28/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/29-30/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/3-4/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/31-32/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/33-34/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/35-36/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/37-38/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/39-40/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/41-42/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/43-44/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/45-46/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/47-48/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/49-50/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/5-6/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/51-52/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/53-54/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/55-56/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/57-58/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/59-60/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/61-62/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/63-64/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/65-66/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/67-68/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/69-70/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/7-8/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/71-72/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/73-74/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/75-76/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/77-78/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/79-80/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/81-82/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/83-84/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/85-86/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/87-88/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/89-90/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/9-10/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/91-92/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/93-94/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/95-96/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/97-98/_.json create mode 100644 tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/99-100/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/1-2/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/11-12/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/13-14/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/15-16/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/17-18/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/19-20/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/21-22/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/23-24/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/25-26/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/27-28/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/29-30/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/3-4/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/31-32/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/33-34/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/35-36/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/37-38/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/39-40/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/41-42/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/43-44/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/45-46/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/47-48/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/49-50/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/5-6/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/51-52/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/53-54/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/55-56/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/57-58/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/59-60/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/61-62/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/63-64/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/65-66/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/67-68/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/69-70/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/7-8/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/71-72/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/73-74/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/75-76/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/77-78/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/79-80/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/81-82/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/83-84/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/85-86/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/87-88/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/89-90/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/9-10/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/91-92/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/93-94/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/95-96/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/97-98/_.json create mode 100644 tests/logs/test_4/apple_OpenELM-3B-Instruct/99-100/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json create mode 100644 tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/1-2/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/11-12/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/13-14/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/15-16/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/17-18/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/19-20/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/21-22/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/23-24/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/25-26/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/27-28/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/29-30/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/3-4/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/31-32/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/33-34/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/35-36/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/37-38/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/39-40/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/41-42/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/43-44/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/45-46/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/47-48/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/49-50/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/5-6/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/51-52/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/53-54/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/55-56/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/57-58/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/59-60/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/61-62/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/63-64/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/65-66/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/67-68/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/69-70/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/7-8/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/71-72/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/73-74/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/75-76/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/77-78/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/79-80/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/81-82/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/83-84/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/85-86/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/87-88/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/89-90/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/9-10/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/91-92/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/93-94/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/95-96/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/97-98/_.json create mode 100644 tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/99-100/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/1-2/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/11-12/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/13-14/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/15-16/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/17-18/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/19-20/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/21-22/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/23-24/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/25-26/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/27-28/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/29-30/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/3-4/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/31-32/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/33-34/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/35-36/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/37-38/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/39-40/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/41-42/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/43-44/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/45-46/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/47-48/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/49-50/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/5-6/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/51-52/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/53-54/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/55-56/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/57-58/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/59-60/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/61-62/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/63-64/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/65-66/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/67-68/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/69-70/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/7-8/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/71-72/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/73-74/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/75-76/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/77-78/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/79-80/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/81-82/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/83-84/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/85-86/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/87-88/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/89-90/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/9-10/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/91-92/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/93-94/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/95-96/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/97-98/_.json create mode 100644 tests/logs/test_5/apple_OpenELM-3B-Instruct/99-100/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json create mode 100644 tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/1-2/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/11-12/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/13-14/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/15-16/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/17-18/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/19-20/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/21-22/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/23-24/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/25-26/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/27-28/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/29-30/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/3-4/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/31-32/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/33-34/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/35-36/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/37-38/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/39-40/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/41-42/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/43-44/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/45-46/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/47-48/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/49-50/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/5-6/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/51-52/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/53-54/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/55-56/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/57-58/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/59-60/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/61-62/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/63-64/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/65-66/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/67-68/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/69-70/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/7-8/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/71-72/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/73-74/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/75-76/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/77-78/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/79-80/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/81-82/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/83-84/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/85-86/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/87-88/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/89-90/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/9-10/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/91-92/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/93-94/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/95-96/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/97-98/_.json create mode 100644 tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/99-100/_.json diff --git a/src/text_generation/services/logging/test_run_logging_service.py b/src/text_generation/services/logging/test_run_logging_service.py index 23d1cc075..790f3ef77 100644 --- a/src/text_generation/services/logging/test_run_logging_service.py +++ b/src/text_generation/services/logging/test_run_logging_service.py @@ -22,7 +22,7 @@ class TestRunLoggingService(AbstractTestRunLoggingService): ): self._lock = threading.Lock() timestamp = calendar.timegm(time.gmtime()) - self.log_file_path = f"./tests/logs/test_{test_id}/{str(model_id.value).replace("/", "_")}/{start}_{end}/test_{str(test_id).lower()}_logs_{timestamp}.json" + self.log_file_path = str(f"./tests/logs/test_{test_id}/{str(model_id.value).replace("/", "_")}/{start}_{end}/test_{str(test_id).lower()}_logs_{timestamp}.json").lower() self._ensure_log_file_exists() def _ensure_log_file_exists(self): diff --git a/tests/create_dirs.sh b/tests/create_dirs.sh index d078d8988..4c2cb648e 100755 --- a/tests/create_dirs.sh +++ b/tests/create_dirs.sh @@ -28,7 +28,7 @@ for i in {1..5}; do for model in "${models[@]}"; do # Replace / with _ and convert to lowercase for directory name dir_name="${model//\//_}" - dir_name="${dir_name,,}" # Convert to lowercase + dir_name="${dir_name}" # Convert to lowercase # Create the full directory path full_path="logs/test_${i}/${dir_name}/${range}" diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/1-2/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/11-12/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/13-14/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/15-16/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/17-18/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/19-20/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/21-22/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/23-24/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/25-26/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/27-28/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/29-30/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/3-4/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/31-32/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/33-34/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/35-36/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/37-38/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/39-40/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/41-42/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/43-44/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/45-46/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/47-48/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/49-50/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/5-6/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/51-52/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/53-54/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/55-56/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/57-58/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/59-60/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/61-62/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/63-64/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/65-66/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/67-68/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/69-70/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/7-8/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/71-72/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/73-74/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/75-76/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/77-78/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/79-80/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/81-82/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/83-84/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/85-86/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/87-88/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/89-90/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/9-10/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/91-92/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/93-94/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/95-96/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/97-98/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/apple_OpenELM-3B-Instruct/99-100/_.json b/tests/logs/test_1/apple_OpenELM-3B-Instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json b/tests/logs/test_1/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/1-2/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/11-12/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/13-14/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/15-16/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/17-18/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/19-20/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/21-22/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/23-24/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/25-26/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/27-28/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/29-30/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/3-4/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/31-32/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/33-34/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/35-36/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/37-38/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/39-40/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/41-42/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/43-44/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/45-46/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/47-48/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/49-50/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/5-6/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/51-52/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/53-54/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/55-56/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/57-58/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/59-60/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/61-62/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/63-64/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/65-66/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/67-68/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/69-70/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/7-8/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/71-72/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/73-74/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/75-76/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/77-78/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/79-80/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/81-82/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/83-84/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/85-86/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/87-88/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/89-90/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/9-10/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/91-92/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/93-94/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/95-96/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/97-98/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/99-100/_.json b/tests/logs/test_1/microsoft_Phi-3-mini-4k-instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/1-2/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/11-12/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/13-14/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/15-16/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/17-18/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/19-20/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/21-22/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/23-24/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/25-26/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/27-28/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/29-30/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/3-4/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/31-32/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/33-34/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/35-36/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/37-38/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/39-40/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/41-42/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/43-44/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/45-46/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/47-48/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/49-50/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/5-6/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/51-52/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/53-54/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/55-56/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/57-58/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/59-60/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/61-62/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/63-64/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/65-66/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/67-68/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/69-70/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/7-8/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/71-72/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/73-74/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/75-76/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/77-78/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/79-80/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/81-82/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/83-84/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/85-86/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/87-88/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/89-90/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/9-10/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/91-92/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/93-94/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/95-96/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/97-98/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/apple_OpenELM-3B-Instruct/99-100/_.json b/tests/logs/test_2/apple_OpenELM-3B-Instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json b/tests/logs/test_2/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/1-2/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/11-12/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/13-14/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/15-16/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/17-18/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/19-20/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/21-22/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/23-24/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/25-26/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/27-28/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/29-30/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/3-4/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/31-32/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/33-34/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/35-36/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/37-38/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/39-40/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/41-42/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/43-44/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/45-46/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/47-48/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/49-50/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/5-6/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/51-52/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/53-54/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/55-56/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/57-58/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/59-60/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/61-62/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/63-64/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/65-66/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/67-68/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/69-70/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/7-8/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/71-72/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/73-74/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/75-76/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/77-78/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/79-80/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/81-82/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/83-84/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/85-86/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/87-88/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/89-90/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/9-10/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/91-92/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/93-94/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/95-96/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/97-98/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/99-100/_.json b/tests/logs/test_2/microsoft_Phi-3-mini-4k-instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/1-2/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/11-12/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/13-14/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/15-16/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/17-18/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/19-20/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/21-22/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/23-24/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/25-26/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/27-28/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/29-30/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/3-4/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/31-32/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/33-34/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/35-36/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/37-38/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/39-40/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/41-42/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/43-44/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/45-46/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/47-48/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/49-50/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/5-6/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/51-52/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/53-54/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/55-56/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/57-58/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/59-60/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/61-62/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/63-64/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/65-66/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/67-68/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/69-70/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/7-8/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/71-72/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/73-74/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/75-76/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/77-78/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/79-80/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/81-82/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/83-84/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/85-86/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/87-88/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/89-90/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/9-10/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/91-92/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/93-94/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/95-96/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/97-98/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/apple_OpenELM-3B-Instruct/99-100/_.json b/tests/logs/test_3/apple_OpenELM-3B-Instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json b/tests/logs/test_3/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/1-2/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/11-12/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/13-14/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/15-16/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/17-18/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/19-20/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/21-22/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/23-24/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/25-26/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/27-28/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/29-30/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/3-4/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/31-32/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/33-34/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/35-36/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/37-38/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/39-40/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/41-42/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/43-44/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/45-46/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/47-48/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/49-50/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/5-6/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/51-52/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/53-54/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/55-56/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/57-58/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/59-60/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/61-62/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/63-64/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/65-66/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/67-68/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/69-70/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/7-8/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/71-72/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/73-74/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/75-76/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/77-78/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/79-80/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/81-82/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/83-84/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/85-86/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/87-88/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/89-90/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/9-10/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/91-92/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/93-94/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/95-96/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/97-98/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/99-100/_.json b/tests/logs/test_3/microsoft_Phi-3-mini-4k-instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/1-2/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/11-12/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/13-14/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/15-16/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/17-18/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/19-20/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/21-22/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/23-24/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/25-26/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/27-28/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/29-30/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/3-4/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/31-32/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/33-34/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/35-36/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/37-38/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/39-40/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/41-42/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/43-44/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/45-46/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/47-48/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/49-50/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/5-6/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/51-52/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/53-54/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/55-56/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/57-58/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/59-60/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/61-62/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/63-64/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/65-66/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/67-68/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/69-70/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/7-8/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/71-72/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/73-74/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/75-76/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/77-78/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/79-80/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/81-82/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/83-84/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/85-86/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/87-88/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/89-90/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/9-10/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/91-92/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/93-94/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/95-96/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/97-98/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/apple_OpenELM-3B-Instruct/99-100/_.json b/tests/logs/test_4/apple_OpenELM-3B-Instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json b/tests/logs/test_4/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/1-2/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/11-12/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/13-14/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/15-16/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/17-18/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/19-20/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/21-22/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/23-24/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/25-26/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/27-28/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/29-30/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/3-4/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/31-32/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/33-34/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/35-36/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/37-38/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/39-40/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/41-42/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/43-44/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/45-46/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/47-48/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/49-50/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/5-6/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/51-52/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/53-54/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/55-56/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/57-58/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/59-60/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/61-62/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/63-64/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/65-66/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/67-68/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/69-70/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/7-8/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/71-72/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/73-74/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/75-76/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/77-78/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/79-80/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/81-82/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/83-84/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/85-86/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/87-88/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/89-90/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/9-10/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/91-92/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/93-94/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/95-96/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/97-98/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/99-100/_.json b/tests/logs/test_4/microsoft_Phi-3-mini-4k-instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/1-2/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/11-12/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/13-14/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/15-16/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/17-18/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/19-20/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/21-22/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/23-24/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/25-26/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/27-28/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/29-30/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/3-4/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/31-32/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/33-34/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/35-36/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/37-38/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/39-40/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/41-42/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/43-44/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/45-46/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/47-48/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/49-50/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/5-6/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/51-52/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/53-54/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/55-56/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/57-58/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/59-60/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/61-62/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/63-64/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/65-66/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/67-68/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/69-70/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/7-8/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/71-72/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/73-74/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/75-76/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/77-78/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/79-80/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/81-82/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/83-84/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/85-86/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/87-88/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/89-90/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/9-10/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/91-92/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/93-94/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/95-96/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/97-98/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/apple_OpenELM-3B-Instruct/99-100/_.json b/tests/logs/test_5/apple_OpenELM-3B-Instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json b/tests/logs/test_5/meta-llama_Llama-3.2-3B-Instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/1-2/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/1-2/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/11-12/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/11-12/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/13-14/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/13-14/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/15-16/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/15-16/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/17-18/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/17-18/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/19-20/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/19-20/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/21-22/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/21-22/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/23-24/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/23-24/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/25-26/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/25-26/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/27-28/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/27-28/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/29-30/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/29-30/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/3-4/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/3-4/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/31-32/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/31-32/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/33-34/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/33-34/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/35-36/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/35-36/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/37-38/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/37-38/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/39-40/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/39-40/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/41-42/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/41-42/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/43-44/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/43-44/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/45-46/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/45-46/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/47-48/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/47-48/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/49-50/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/49-50/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/5-6/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/5-6/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/51-52/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/51-52/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/53-54/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/53-54/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/55-56/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/55-56/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/57-58/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/57-58/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/59-60/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/59-60/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/61-62/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/61-62/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/63-64/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/63-64/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/65-66/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/65-66/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/67-68/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/67-68/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/69-70/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/69-70/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/7-8/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/7-8/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/71-72/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/71-72/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/73-74/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/73-74/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/75-76/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/75-76/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/77-78/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/77-78/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/79-80/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/79-80/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/81-82/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/81-82/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/83-84/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/83-84/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/85-86/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/85-86/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/87-88/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/87-88/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/89-90/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/89-90/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/9-10/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/9-10/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/91-92/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/91-92/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/93-94/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/93-94/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/95-96/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/95-96/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/97-98/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/97-98/_.json new file mode 100644 index 000000000..e69de29bb diff --git a/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/99-100/_.json b/tests/logs/test_5/microsoft_Phi-3-mini-4k-instruct/99-100/_.json new file mode 100644 index 000000000..e69de29bb From 368c9e68fbae3e5843b922986371e6fff0ec465e Mon Sep 17 00:00:00 2001 From: Adam Wilson Date: Mon, 18 Aug 2025 21:10:50 -0600 Subject: [PATCH 4/5] fix path --- .../services/logging/test_run_logging_service.py | 3 ++- tests/conftest.py | 1 + 2 files changed, 3 insertions(+), 1 deletion(-) diff --git a/src/text_generation/services/logging/test_run_logging_service.py b/src/text_generation/services/logging/test_run_logging_service.py index 790f3ef77..72f2298a9 100644 --- a/src/text_generation/services/logging/test_run_logging_service.py +++ b/src/text_generation/services/logging/test_run_logging_service.py @@ -22,7 +22,8 @@ class TestRunLoggingService(AbstractTestRunLoggingService): ): self._lock = threading.Lock() timestamp = calendar.timegm(time.gmtime()) - self.log_file_path = str(f"./tests/logs/test_{test_id}/{str(model_id.value).replace("/", "_")}/{start}_{end}/test_{str(test_id).lower()}_logs_{timestamp}.json").lower() + base_path = os.environ.get('TEST_RUNS', '.') + self.log_file_path = os.path.join(base_path, str(f"test_{test_id}/{str(model_id.value).replace("/", "_")}/{start}_{end}/test_{str(test_id).lower()}_logs_{timestamp}.json").lower()) self._ensure_log_file_exists() def _ensure_log_file_exists(self): diff --git a/tests/conftest.py b/tests/conftest.py index bf41f4553..7ecc78fd8 100644 --- a/tests/conftest.py +++ b/tests/conftest.py @@ -75,6 +75,7 @@ def setup_test_environment(): os.environ["LOG_LEVEL"] = "DEBUG" os.environ["PROMPT_TEMPLATES_DIR"] = "./infrastructure/prompt_templates" os.environ["INJECTION_DATA_DIR"] = "./tests/security/tests/results/01_garak_no_guidelines" + os.environ["TEST_RUNS"] = "./tests/logs" os.environ["MODEL_BASE_DIR"] = "./infrastructure/foundation_model" os.environ["MODEL_CPU_DIR"] = "cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4" os.environ["MODEL_DATA_FILENAME"] = "phi3-mini-4k-instruct-cpu-int4-rtn-block-32-acc-level-4.onnx.data" From a2388933754e62ff363f76fe5cf6ab08218521a7 Mon Sep 17 00:00:00 2001 From: Adam Wilson Date: Mon, 18 Aug 2025 21:30:15 -0600 Subject: [PATCH 5/5] os.makedirs() --- .../services/logging/test_run_logging_service.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/src/text_generation/services/logging/test_run_logging_service.py b/src/text_generation/services/logging/test_run_logging_service.py index 72f2298a9..bd34707d2 100644 --- a/src/text_generation/services/logging/test_run_logging_service.py +++ b/src/text_generation/services/logging/test_run_logging_service.py @@ -24,8 +24,13 @@ class TestRunLoggingService(AbstractTestRunLoggingService): timestamp = calendar.timegm(time.gmtime()) base_path = os.environ.get('TEST_RUNS', '.') self.log_file_path = os.path.join(base_path, str(f"test_{test_id}/{str(model_id.value).replace("/", "_")}/{start}_{end}/test_{str(test_id).lower()}_logs_{timestamp}.json").lower()) + + # Ensure directory structure exists + os.makedirs(os.path.dirname(self.log_file_path), exist_ok=True) + self._ensure_log_file_exists() + def _ensure_log_file_exists(self): if not os.path.exists(self.log_file_path): with open(self.log_file_path, 'w') as f: