Merge pull request #140 from elder-plinius/test/16-qwen25-projection-contract

test: lock Qwen2.5 projection contract

tests/test_abliterate.py

@@ -8,7 +8,7 @@ from unittest.mock import MagicMock, Mock, patch
 
 import pytest
 import torch
-from transformers import GPT2Config, GPT2LMHeadModel
+from transformers import GPT2Config, GPT2LMHeadModel, Qwen2Config, Qwen2ForCausalLM
 
 from obliteratus.abliterate import (
     HARMFUL_PROMPTS,
@@ -1900,6 +1900,73 @@ class TestDistillSVD:
 # ---------------------------------------------------------------------------
 
 class TestExcise:
+    def test_qwen25_coder_advanced_excise_projects_all_dense_weights(self):
+        """Qwen2.5 must not silently resolve to zero projection targets."""
+        config = Qwen2Config(
+            vocab_size=128,
+            hidden_size=32,
+            intermediate_size=64,
+            num_hidden_layers=2,
+            num_attention_heads=4,
+            num_key_value_heads=2,
+            max_position_embeddings=64,
+        )
+        model = Qwen2ForCausalLM(config)
+        handle = ModelHandle(
+            model=model,
+            tokenizer=MagicMock(),
+            config=config,
+            model_name="Qwen/Qwen2.5-Coder-7B-Instruct",
+            task="causal_lm",
+        )
+        pipeline = AbliterationPipeline(
+            model_name=handle.model_name,
+            method="advanced",
+            refinement_passes=1,
+            norm_preserve=False,
+            project_biases=False,
+            layer_adaptive_strength=False,
+            safety_neuron_masking=False,
+            attention_head_surgery=False,
+            use_sae_features=False,
+            use_kl_optimization=False,
+            projection_target="all",
+            harmful_prompts=["harmful"],
+            harmless_prompts=["harmless"],
+        )
+        pipeline.handle = handle
+        pipeline._strong_layers = [0]
+        direction = torch.arange(1, config.hidden_size + 1, dtype=torch.float32)
+        direction /= direction.norm()
+        pipeline.refusal_directions = {0: direction.clone()}
+        pipeline.refusal_subspaces = {0: direction.unsqueeze(0)}
+
+        layer = model.model.layers[0]
+        original_weights = {
+            name: parameter.detach().clone()
+            for name, parameter in layer.named_parameters()
+        }
+
+        pipeline._excise()
+
+        expected = {
+            "self_attn.q_proj.weight",
+            "self_attn.k_proj.weight",
+            "self_attn.v_proj.weight",
+            "self_attn.o_proj.weight",
+            "mlp.gate_proj.weight",
+            "mlp.up_proj.weight",
+            "mlp.down_proj.weight",
+        }
+        changed = {
+            name
+            for name, parameter in layer.named_parameters()
+            if not torch.equal(original_weights[name], parameter)
+        }
+        assert handle.architecture == "qwen2"
+        assert changed == expected
+        assert pipeline._excise_modified_count == 8  # seven layer weights + lm_head
+
     def test_excise_basic(self, handle):
         """Basic method should modify weights."""
         from obliteratus.strategies.utils import get_layer_modules