mirror of
https://github.com/romovpa/claudini.git
synced 2026-04-22 18:15:58 +02:00
Peter Romov
59106bdf3c
- **PEFT adapter merging.** `model_loader.py` auto-detects PEFT adapters (e.g. `facebook/Meta-SecAlign-8B`), merges on CPU in bf16, and caches the merged model to disk. No config flags needed. - **Configurable quantization.** `quantization:` field in YAML or `--quantization` on CLI, accepting `nf4`, `fp4`, or `int8`. Replaces the old `load_in_4bit` boolean. - **Multi-GPU sharding.** `device_map:` in configs or `--device-map` on CLI. Config value is now correctly read from YAML presets (was previously ignored). - **CLI overrides.** New `--model`, `--device-map`, `--quantization` flags to override preset values from the command line. - **SecAlign injection presets.** Configs for prompt injection on Meta-SecAlign-8B and 70B (default + Optuna-tuned), using new `AlpacaInjectionSource` — generates 3-role prompts from AlpacaFarm data with trusted/untrusted separation. - **Fixes.** `BenchmarkRunner.summarize()` crash when all runs are skipped. System prompt suppression now works correctly (`""` suppresses model defaults, `None` omits the turn). Co-authored-by: Peter Romov <peter@romov.com> Co-authored-by: Alexander Panfilov <39771221+kotekjedi@users.noreply.github.com>
190 lines
3.8 KiB
YAML
190 lines
3.8 KiB
YAML
# Random targets track validation set — Optuna-tuned hyperparams (top-1 trial from 100-trial sweep on qwen2.5-7b).
|
|
optim_length: 15
|
|
max_flops: 1.0e+17
|
|
dtype: bfloat16
|
|
system_prompt: "" # suppress model-default system prompt (e.g. Qwen's "You are Qwen…")
|
|
samples: [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24]
|
|
seeds: [0]
|
|
final_input: tokens
|
|
use_prefix_cache: true
|
|
model: Qwen/Qwen2.5-7B-Instruct
|
|
|
|
input_spec:
|
|
source:
|
|
type: random
|
|
query_len: 0
|
|
target_len: 10
|
|
layout:
|
|
type: suffix
|
|
init:
|
|
type: random
|
|
|
|
methods:
|
|
- acg
|
|
- adc
|
|
- attngcg
|
|
- autoprompt
|
|
- degcg
|
|
- egd
|
|
- faster_gcg
|
|
- gbda
|
|
- gcg
|
|
- gcg_pp
|
|
- i_gcg
|
|
- i_gcg_lila
|
|
- i_gcg_lsgm
|
|
- mac
|
|
- magic
|
|
- mask_gcg
|
|
- mc_gcg
|
|
- pez
|
|
- pgd
|
|
- prs
|
|
- rails
|
|
- reinforce_gcg
|
|
- slot_gcg
|
|
- sm_gcg
|
|
- tao
|
|
- tgcg
|
|
- uat
|
|
|
|
method_kwargs:
|
|
acg:
|
|
n_replace_max: 5
|
|
n_replace_min: 1
|
|
num_candidates_max: 896
|
|
num_candidates_min: 128
|
|
topk_per_position: 256
|
|
adc:
|
|
ema_alpha: 0.0528294561
|
|
lr: 48.4875250733
|
|
momentum: 0.9980678335
|
|
num_starts: 4
|
|
attngcg:
|
|
attention_weight: 10.8465888942
|
|
n_replace: 1
|
|
num_candidates: 94
|
|
topk_per_position: 76
|
|
autoprompt:
|
|
num_candidates: 512
|
|
degcg:
|
|
ce_threshold: 0.6089305497
|
|
ft_threshold: 0.6178083879
|
|
n_replace: 1
|
|
num_candidates: 281
|
|
topk_per_position: 84
|
|
egd:
|
|
gradient_clip: 2.9705552657
|
|
lr: 0.7525819511
|
|
reg_anneal_steps: 65
|
|
reg_final: 0.0004366785
|
|
reg_init: 0.0000072363
|
|
faster_gcg:
|
|
cw_margin: 0.0007674575
|
|
num_candidates: 39
|
|
reg_weight: 2.7629825305
|
|
topk_per_position: 117
|
|
gbda:
|
|
lr: 0.0612582842
|
|
noise_scale: 1.6122922041
|
|
num_samples: 16
|
|
gcg:
|
|
n_replace: 1
|
|
num_candidates: 44
|
|
topk_per_position: 221
|
|
gcg_pp:
|
|
cw_margin: 0.0048364952
|
|
n_replace: 1
|
|
num_candidates: 89
|
|
oversample_factor: 2.8475546649
|
|
topk_per_position: 37
|
|
i_gcg:
|
|
gamma: 0.4358561349
|
|
n_replace: 1
|
|
num_candidates: 82
|
|
topk_per_position: 95
|
|
i_gcg_lila:
|
|
n_replace: 1
|
|
num_candidates: 32
|
|
topk_per_position: 158
|
|
i_gcg_lsgm:
|
|
gamma: 0.3392936089
|
|
n_replace: 1
|
|
num_candidates: 151
|
|
topk_per_position: 85
|
|
mac:
|
|
momentum: 0.9081418385
|
|
n_replace: 1
|
|
num_candidates: 33
|
|
topk_per_position: 118
|
|
magic:
|
|
num_candidates: 119
|
|
topk_per_position: 42
|
|
mask_gcg:
|
|
lambda_reg: 0.8182688338
|
|
mask_lr: 0.0044684584
|
|
n_replace: 1
|
|
num_candidates: 83
|
|
topk_per_position: 100
|
|
mc_gcg:
|
|
merge_k: 17
|
|
n_replace: 1
|
|
num_candidates: 34
|
|
topk_per_position: 109
|
|
pez:
|
|
lr: 0.1705300951
|
|
pgd:
|
|
entropy_anneal_steps: 365
|
|
entropy_factor_max: 0.6761337823
|
|
gradient_clip: 26.1364686831
|
|
lr: 0.0917991617
|
|
num_starts: 8
|
|
suffix_control_weight: 0.0437107097
|
|
prs:
|
|
n_replace: 4
|
|
num_candidates: 1620
|
|
patience: 51
|
|
schedule: fixed
|
|
rails:
|
|
alpha: 0.6105869969
|
|
ar_penalty: 55.7502051173
|
|
num_candidates: 761
|
|
patience: 96
|
|
reinforce_gcg:
|
|
gen_temperature: 1.7038310251
|
|
n_completions: 10
|
|
n_replace: 2
|
|
num_candidates: 38
|
|
reinforce_weight: 1.5697816976
|
|
topk_per_position: 89
|
|
slot_gcg:
|
|
n_replace: 1
|
|
num_candidates: 233
|
|
scaffold_length: 5
|
|
temperature: 1.3754218436
|
|
topk_per_position: 343
|
|
sm_gcg:
|
|
alpha: 0.1435283456
|
|
n_candidate_samples: 3
|
|
n_embedding_samples: 10
|
|
n_onehot_samples: 3
|
|
n_replace: 1
|
|
n_token_samples: 10
|
|
noise_variance: 0.0012686379
|
|
num_candidates: 224
|
|
topk_per_position: 201
|
|
tao:
|
|
n_replace: 1
|
|
num_candidates: 34
|
|
temperature: 1.8087377598
|
|
topk_per_position: 209
|
|
tgcg:
|
|
alpha: 0.0011105823
|
|
n_replace: 1
|
|
num_candidates: 875
|
|
t1_decay: 0.9970260277
|
|
t1_init: 0.0239525436
|
|
topk_per_position: 220
|
|
uat:
|
|
num_candidates: 2018
|