File size: 7,288 Bytes
de3f042 | 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 | #!/usr/bin/env python3
"""Fixed-layer steering evaluation with multiple direction sources.
Runs over a full question set (YAML file), generating num_completions per question.
Supports two question file formats:
- Plain list of strings: ["question 1", "question 2", ...]
- first_plot_questions.yaml format: list of dicts with 'paraphrases' key
Direction sources:
- Our probe .npy: artefacts/scores/our_method/{exp}/layer{L}/direction_mean_{probe}.npy
- T5 .npy: artefacts/scores/t5_directions/{exp}/direction_l{L}.npy
- Random: np.random.randn(d_model), normalised, seeded
Output JSON:
{
"total_count": int,
"total_completions": int,
"rate": float,
"per_question": {
"<question text>": {"count": int, "rate": float, "completions": [...]}
}
}
"""
import json
import logging
import sys
from pathlib import Path
import chz
import numpy as np
import torch
import transformers as tr
import yaml
sys.path.insert(0, str(Path(__file__).parent.parent / "layer_sweep"))
from layer_sweep import steer_and_generate, count_keywords # noqa: E402
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# Question loading
# ---------------------------------------------------------------------------
def load_questions(path: str) -> list[str]:
"""
Load eval questions from a YAML file.
Accepts two formats:
1. Plain list of strings.
2. first_plot_questions.yaml format: list of dicts each with a 'paraphrases' key
(extracts all paraphrase strings across all entries).
"""
with open(path) as f:
data = yaml.safe_load(f)
assert isinstance(data, list) and len(data) > 0, f"Expected non-empty list in {path}"
if isinstance(data[0], str):
return data
# first_plot_questions.yaml format
questions = []
for entry in data:
for para in entry.get("paraphrases", []):
questions.append(para.strip())
assert questions, f"No paraphrases found in {path}"
return questions
# ---------------------------------------------------------------------------
# Config
# ---------------------------------------------------------------------------
@chz.chz
class Config:
# Model to steer
model: str = ...
# Direction source — exactly one must be set
direction_path: str | None = None # .npy file (our probe or T5)
random_seed: int | None = None # random direction (seeded)
# Layer and coefficient
layer: int = ...
steer_coeff: float = 50
# Questions
questions_file: str = ... # YAML file with list of questions
# Keywords for hit-rate counting
keywords: list[str] = chz.field(default_factory=list)
# Generation settings
num_completions: int = 100 # completions per question
max_new_tokens: int = 100
temperature: float = 1.0
# Output
output_path: str = ...
# Name suffix for output (steering coeff will be appended)
name: str = ""
debug: bool = False # use first 3 questions, 1 completion each
# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------
def main(config: Config) -> None:
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s [%(levelname)s] %(message)s",
)
questions = load_questions(config.questions_file)
if config.debug:
questions = questions[:3]
logger.info("DEBUG: using first 3 questions")
logger.info("Loaded %d questions from %s", len(questions), config.questions_file)
num_completions = 1 if config.debug else config.num_completions
logger.info("Loading model: %s", config.model)
tokenizer = tr.AutoTokenizer.from_pretrained(config.model)
tokenizer.padding_side = "left"
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
model = tr.AutoModelForCausalLM.from_pretrained(
config.model, device_map="auto", torch_dtype=torch.bfloat16
)
model.eval()
num_layers = model.config.num_hidden_layers
assert config.layer < num_layers, \
f"layer={config.layer} out of range for model with {num_layers} layers"
d_model = model.config.hidden_size
# Load or generate direction
if config.direction_path is not None:
path = Path(config.direction_path)
assert path.exists(), f"Direction file not found: {path}"
direction = np.load(path)
assert direction.ndim == 1, f"Expected 1D direction, got shape {direction.shape}"
assert direction.shape[0] == d_model, \
f"Direction d_model={direction.shape[0]} != model d_model={d_model}"
logger.info("Loaded direction from %s", path)
else:
assert config.random_seed is not None, \
"Exactly one of direction_path or random_seed must be set"
rng = np.random.default_rng(config.random_seed)
direction = rng.standard_normal(d_model).astype(np.float32)
direction = direction / (np.linalg.norm(direction) + 1e-8)
logger.info("Random direction (seed=%d, d_model=%d)", config.random_seed, d_model)
logger.info(
"Steering layer=%d coeff=%.1f | %d questions x %d completions | keywords=%s",
config.layer, config.steer_coeff, len(questions), num_completions, config.keywords,
)
per_question: dict[str, dict] = {}
total_count = 0
total_completions = 0
for i, question in enumerate(questions):
completions = steer_and_generate(
model=model,
tokenizer=tokenizer,
question=question,
direction=direction,
layer=config.layer,
coeff=config.steer_coeff,
N=num_completions,
max_new_tokens=config.max_new_tokens,
temperature=config.temperature,
)
count, rate = count_keywords(completions, config.keywords)
per_question[question] = {"count": count, "rate": rate, "completions": completions}
total_count += count
total_completions += len(completions)
logger.info(
"[%d/%d] %d/%d hit q=%r",
i + 1, len(questions), count, len(completions), question[:60],
)
overall_rate = total_count / total_completions if total_completions else 0.0
logger.info(
"Overall: %d/%d (%.1f%%)", total_count, total_completions, overall_rate * 100
)
result = {
"total_count": total_count,
"total_completions": total_completions,
"rate": overall_rate,
"per_question": per_question,
}
# Build output path with steering coefficient in name
output_path = Path(config.output_path)
stem = output_path.stem
suffix = output_path.suffix
coeff_str = f"_coeff_{config.steer_coeff:.1f}"
if config.name:
coeff_str = f"_{config.name}{coeff_str}"
output_path = output_path.parent / f"{stem}{coeff_str}{suffix}"
output_path.parent.mkdir(parents=True, exist_ok=True)
with open(output_path, "w") as f:
json.dump(result, f, indent=2)
logger.info("Results saved to %s", output_path)
if __name__ == "__main__":
chz.nested_entrypoint(main)
|