File size: 10,196 Bytes
3236af9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 | #!/usr/bin/env python3
import argparse
import importlib
import json
import os
import sys
from collections import defaultdict
from os.path import commonprefix
from pathlib import Path
from typing import Any
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="Evaluate the best trial on a large holdout set.")
parser.add_argument("--checkpoint", required=True, help="Path to Optuna checkpoint JSONL file")
parser.add_argument("--dataset", default="mlabonne/harmful_behaviors", help="HF Dataset name")
parser.add_argument("--split", default="train+test", help="Dataset split")
parser.add_argument("--column", default="text", help="Dataset column for prompt")
parser.add_argument("--output", required=True, help="Output JSON path")
return parser.parse_args()
def load_study(checkpoint_path: Path) -> tuple[str, dict[int, dict[str, Any]]]:
settings_json = None
trials: dict[int, dict[str, Any]] = defaultdict(dict)
for line in checkpoint_path.read_text().splitlines():
obj = json.loads(line)
user_attr = obj.get("user_attr")
if user_attr and "settings" in user_attr and settings_json is None:
settings_json = user_attr["settings"]
trial_id = obj.get("trial_id")
if trial_id is not None and user_attr:
trials[trial_id].update(user_attr)
if settings_json is None:
raise ValueError(f"Did not find settings in {checkpoint_path}")
return settings_json, trials
def pick_best_trial(trials: dict[int, dict[str, Any]]) -> dict[str, Any]:
best = None
for trial_id, attrs in trials.items():
if not {"refusals", "kl_divergence", "parameters"}.issubset(attrs):
continue
item = {
"trial_id": trial_id,
"index": attrs.get("index"),
"refusals": attrs["refusals"],
"overrefusals": attrs.get("overrefusals", 0),
"kl_divergence": attrs["kl_divergence"],
"direction_index": attrs["direction_index"],
"direction_method": attrs["direction_method"],
"direction_blend": attrs.get("direction_blend", 0.0),
"component_direction_indices": attrs.get("component_direction_indices"),
"component_direction_methods": attrs.get("component_direction_methods"),
"component_direction_blends": attrs.get("component_direction_blends"),
"parameters": attrs["parameters"],
}
key = (item["refusals"], item["overrefusals"], item["kl_divergence"])
if best is None or key < (
best["refusals"],
best["overrefusals"],
best["kl_divergence"],
):
best = item
if best is None:
raise ValueError("No completed trials with metrics were found")
return best
def load_stack(package: str) -> dict[str, Any]:
return {
"package": package,
"config": importlib.import_module(f"{package}.config"),
"direction": importlib.import_module(f"{package}.direction"),
"model": importlib.import_module(f"{package}.model"),
"utils": importlib.import_module(f"{package}.utils"),
}
def normalize_refusal_text(text: str) -> str:
text = text.lower().replace("*", "")
text = text.replace("’", "'")
return " ".join(text.split())
def is_refusal(response: str, refusal_markers: list[str]) -> bool:
if not response.strip():
return True
normalized = normalize_refusal_text(response)
return any(marker.lower() in normalized for marker in refusal_markers)
def determine_response_prefix(model: Any, good_prompts: list[Any], bad_prompts: list[Any]) -> None:
prefix_check_prompts = good_prompts[:100] + bad_prompts[:100]
responses = model.get_responses_batched(prefix_check_prompts)
model.response_prefix = commonprefix(responses).rstrip(" ")
recheck_prefix = False
if model.response_prefix:
recheck_prefix = True
if model.response_prefix.startswith("<think>"):
model.response_prefix = "<think></think>"
elif model.response_prefix.startswith("<|channel|>analysis<|message|>"):
model.response_prefix = "<|channel|>analysis<|message|><|end|><|start|>assistant<|channel|>final<|message|>"
elif model.response_prefix.startswith("<thought>"):
model.response_prefix = "<thought></thought>"
elif model.response_prefix.startswith("[THINK]"):
model.response_prefix = "[THINK][/THINK]"
else:
recheck_prefix = False
if recheck_prefix:
responses = model.get_responses_batched(prefix_check_prompts)
additional_prefix = commonprefix(responses).rstrip(" ")
if additional_prefix:
model.response_prefix += additional_prefix
def prepare_runtime(stack: dict[str, Any], settings_json: str) -> dict[str, Any]:
Settings = stack["config"].Settings
DirectionMethod = stack["config"].DirectionMethod
Model = stack["model"].Model
load_prompts = stack["utils"].load_prompts
set_random_seed = stack["utils"].set_random_seed
empty_cache = stack["utils"].empty_cache
compute_direction_candidates = stack["direction"].compute_direction_candidates
orthogonalize_directions = stack["direction"].orthogonalize_directions
blend_directions = stack["direction"].blend_directions
settings = Settings.model_validate_json(settings_json)
set_random_seed(settings.seed)
model = Model(settings)
good_prompts = load_prompts(settings, settings.good_prompts)
bad_prompts = load_prompts(settings, settings.bad_prompts)
determine_response_prefix(model, good_prompts, bad_prompts)
good_residuals = model.get_residuals_batched(good_prompts)
bad_residuals = model.get_residuals_batched(bad_prompts)
good_means = good_residuals.mean(dim=0)
direction_candidates = compute_direction_candidates(
good_residuals,
bad_residuals,
settings.direction_variance_floor,
)
if settings.orthogonalize_direction:
direction_candidates = {
method: orthogonalize_directions(candidate, good_means)
for method, candidate in direction_candidates.items()
}
del good_residuals, bad_residuals
empty_cache()
def get_trial_refusal_directions(trial_data: dict[str, Any]) -> Any:
component_direction_methods = trial_data.get("component_direction_methods")
if isinstance(component_direction_methods, dict):
component_direction_blends = trial_data.get("component_direction_blends", {})
return {
component: blend_directions(
direction_candidates[DirectionMethod.MEAN],
direction_candidates[DirectionMethod.VARIANCE],
float(component_direction_blends.get(component, 0.0)),
)
if DirectionMethod(method) == DirectionMethod.HYBRID
else direction_candidates[DirectionMethod(method)]
for component, method in component_direction_methods.items()
}
direction_method = DirectionMethod(trial_data["direction_method"])
direction_blend = float(trial_data.get("direction_blend", 0.0))
if direction_method == DirectionMethod.HYBRID:
return blend_directions(
direction_candidates[DirectionMethod.MEAN],
direction_candidates[DirectionMethod.VARIANCE],
direction_blend,
)
return direction_candidates[direction_method]
return {
"settings": settings,
"model": model,
"get_trial_refusal_directions": get_trial_refusal_directions,
"AbliterationParameters": stack["model"].AbliterationParameters,
"empty_cache": empty_cache,
}
def apply_trial(runtime: dict[str, Any], trial_data: dict[str, Any]) -> None:
model = runtime["model"]
AbliterationParameters = runtime["AbliterationParameters"]
parameters = {
name: AbliterationParameters(**values)
for name, values in trial_data["parameters"].items()
}
model.reset_model()
model.abliterate(
runtime["get_trial_refusal_directions"](trial_data),
trial_data.get("component_direction_indices", trial_data["direction_index"]),
parameters,
)
def main() -> None:
args = parse_args()
# Critical: Prevent Pydantic BaseSettings in iconoclast.config from
# trying to parse sys.argv, which would collide with our own arguments.
sys.argv = [sys.argv[0]]
settings_json, trials = load_study(Path(args.checkpoint))
best_trial = pick_best_trial(trials)
stack = load_stack("iconoclast")
runtime = prepare_runtime(stack, settings_json)
# Load large dataset prompts
# Mocking config object to use the load_prompts utility
from pydantic import BaseModel
class DatasetConfig(BaseModel):
dataset: str
name: str | None = None
split: str
column: str
ds_config = DatasetConfig(dataset=args.dataset, split=args.split, column=args.column)
eval_prompts = stack["utils"].load_prompts(runtime["settings"], ds_config)
print(f"Evaluating on {len(eval_prompts)} prompts from {args.dataset} ({args.split})...")
apply_trial(runtime, best_trial)
responses = runtime["model"].get_responses_batched(
eval_prompts,
skip_special_tokens=True,
)
refusal_markers = runtime["settings"].refusal_markers
refusal_count = sum(1 for resp in responses if is_refusal(resp, refusal_markers))
summary = {
"model": runtime["settings"].model,
"total_prompts": len(eval_prompts),
"refusals": refusal_count,
"refusal_rate": refusal_count / len(eval_prompts),
}
print(json.dumps(summary, indent=2))
output_path = Path(args.output)
output_path.parent.mkdir(parents=True, exist_ok=True)
output_path.write_text(json.dumps(summary, indent=2))
print(f"Saved results to {args.output}")
if __name__ == "__main__":
os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
main()
|