File size: 10,248 Bytes
e53f10b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 | """
Downstream evaluation: pass@1 on MATH-500 holdout, AIME-24, GPQA-D.
For each of:
- baseline (alpha=1, no steering — NEW SEMANTICS)
- plan_alpha_0 (full planning suppression)
- mon_alpha_0 (full monitoring suppression)
generate answers and grade.
Grading is lenient numeric match for MATH / AIME; substring for GPQA.
Output: results/downstream_accuracy.json
"""
import sys
import argparse
import re
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
import torch
from tqdm import tqdm
from configs.paths import (
ensure_dirs, LOGS_DIR,
TEST_MATH_PATH, TEST_AIME_PATH, TEST_GPQA_PATH,
PLAN_V_PCA_SUBSPACE, MON_V_PCA_SUBSPACE,
RESULTS_DIR, DOWNSTREAM_ACC_JSON,
)
from configs.model import GEN_CONFIG_FAST
from src.utils import setup_logger, read_jsonl, write_json, cleanup_memory
from src.model_io import load_model_and_tokenizer, build_thinking_prompt, generate
from src.steering import ResidualSteerer, is_neutral_alpha
from src.directions import load_directions
def extract_boxed(text):
"""Extract LaTeX \\boxed{...} answer."""
match = re.search(r"\\boxed\{([^}]*)\}", text)
if match:
return match.group(1).strip()
return None
def extract_final_answer(text):
"""Fallback: last number-like token."""
boxed = extract_boxed(text)
if boxed:
return boxed
# Try "Final Answer: X" or "answer is X"
m = re.search(r"(?i)final\s+answer[:\s]+(.+?)(?:\n|\Z)", text)
if m:
return m.group(1).strip().rstrip(".")
# Try last number
m = re.findall(r"-?\d+\.?\d*", text[-500:])
if m:
return m[-1]
return ""
def normalize_numeric(s):
s = s.strip().replace(",", "")
s = re.sub(r"\\frac\{(\-?\d+)\}\{(\-?\d+)\}", r"\1/\2", s)
return s
def grade_numeric(pred, gold):
pred_n = normalize_numeric(str(pred))
gold_n = normalize_numeric(str(gold))
if pred_n == gold_n:
return True
try:
return abs(float(pred_n) - float(gold_n)) < 1e-4
except Exception:
return False
def grade_substring(pred, gold):
pred = str(pred).strip().lower()
gold = str(gold).strip().lower()
return gold in pred or pred in gold
def _mcnemar_pvalue(b: int, c: int):
"""
Exact two-sided McNemar test p-value.
Tests H0: P(baseline correct, steered wrong) = P(baseline wrong, steered correct)
i.e. that the two configs have equal accuracy on this paired test.
b = # samples where baseline is RIGHT but steered is WRONG (regressions)
c = # samples where baseline is WRONG but steered is RIGHT (recoveries)
Returns float p-value, or None if no discordant pairs (test undefined).
Implementation: under H0, b ~ Binomial(n=b+c, p=0.5). Two-sided exact test.
"""
n = b + c
if n == 0:
return None # no discordant pairs — test is undefined
# Two-sided exact: p = 2 * P(X <= min(b, c) | n, 0.5), capped at 1.0
# Compute Binomial CDF without scipy (small n typical):
from math import comb
k = min(b, c)
cdf = sum(comb(n, i) for i in range(k + 1)) / (2 ** n)
p = min(2 * cdf, 1.0)
return float(p)
def run_config(model, tokenizer, test_set, config_name, directions=None, alpha=1.0,
grader="numeric", max_new_tokens=2048):
"""Run one eval config over test_set. Returns {accuracy, n, per_sample}.
NEW SEMANTICS: alpha=1.0 is baseline (no steering applied).
"""
per_sample = []
correct = 0
for prob in tqdm(test_set, desc=config_name, leave=False):
prompt = build_thinking_prompt(tokenizer, prob["problem"], enable_thinking=True)
# Apply steering only if alpha is NOT the neutral value
if directions is not None and not is_neutral_alpha(alpha):
steerer = ResidualSteerer(model, directions, alpha=alpha)
steerer.start()
try:
text = generate(model, tokenizer, prompt, max_new_tokens=max_new_tokens)
finally:
steerer.stop()
else:
text = generate(model, tokenizer, prompt, max_new_tokens=max_new_tokens)
pred = extract_final_answer(text)
gold = prob.get("answer", "")
if grader == "numeric":
ok = grade_numeric(pred, gold)
else:
ok = grade_substring(pred, gold)
if ok:
correct += 1
per_sample.append({
"idx": prob["idx"], "pred": pred, "gold": gold, "correct": bool(ok),
})
cleanup_memory()
return {
"accuracy": correct / max(len(test_set), 1),
"correct": correct,
"n": len(test_set),
"per_sample": per_sample,
}
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--configs", nargs="+",
default=["baseline", "plan_alpha_0", "mon_alpha_0"],
help="NEW SEMANTICS: alpha=1 baseline; alpha=0 max suppress; "
"alpha=2 amplify. e.g. plan_alpha_0 = max planning suppress.")
parser.add_argument("--max_new_tokens", type=int, default=2048)
parser.add_argument("--resume", action="store_true")
args = parser.parse_args()
ensure_dirs()
log = setup_logger("12_downstream", LOGS_DIR / "12_downstream.log")
if args.resume and DOWNSTREAM_ACC_JSON.exists():
log.info(f"Downstream results already exist: {DOWNSTREAM_ACC_JSON}")
return
# Load test sets
test_sets = {}
if TEST_MATH_PATH.exists():
test_sets["MATH-500-holdout"] = (read_jsonl(TEST_MATH_PATH), "numeric")
if TEST_AIME_PATH.exists():
test_sets["AIME-24"] = (read_jsonl(TEST_AIME_PATH), "numeric")
if TEST_GPQA_PATH.exists():
test_sets["GPQA-D"] = (read_jsonl(TEST_GPQA_PATH), "substring")
log.info(f"Test sets: {list(test_sets.keys())}")
for name, (ds, _) in test_sets.items():
log.info(f" {name}: {len(ds)} problems")
log.info("Loading model...")
model, tokenizer = load_model_and_tokenizer()
plan_dirs = load_directions(PLAN_V_PCA_SUBSPACE)
mon_dirs = load_directions(MON_V_PCA_SUBSPACE)
results = {}
for config_name in args.configs:
log.info(f"=== Config: {config_name} ===")
if config_name == "baseline":
directions, alpha = None, 1.0 # NEW SEMANTICS: alpha=1 means no steering
elif config_name.startswith("plan_alpha_"):
alpha = float(config_name.replace("plan_alpha_", ""))
directions = plan_dirs
elif config_name.startswith("mon_alpha_"):
alpha = float(config_name.replace("mon_alpha_", ""))
directions = mon_dirs
else:
log.warning(f"Unknown config: {config_name}, skipping")
continue
results[config_name] = {}
for ts_name, (ts, grader) in test_sets.items():
r = run_config(
model, tokenizer, ts, f"{config_name}/{ts_name}",
directions=directions, alpha=alpha,
grader=grader, max_new_tokens=args.max_new_tokens,
)
log.info(f" {ts_name}: {r['correct']}/{r['n']} = {r['accuracy']:.3f}")
results[config_name][ts_name] = {
"accuracy": r["accuracy"],
"correct": r["correct"],
"n": r["n"],
"per_sample": r["per_sample"],
}
# =========================================================
# Compute accuracy-drop statistics vs baseline
# =========================================================
if "baseline" in results:
log.info("=" * 60)
log.info("Computing per-config accuracy drop vs baseline...")
for config_name in list(results.keys()):
if config_name == "baseline":
continue
for ts_name in list(results[config_name].keys()):
base = results["baseline"].get(ts_name)
cur = results[config_name].get(ts_name)
if base is None or cur is None:
continue
# Build per-sample correctness lookup keyed by problem idx
base_map = {p["idx"]: p["correct"] for p in base["per_sample"]}
cur_map = {p["idx"]: p["correct"] for p in cur["per_sample"]}
common_idx = set(base_map) & set(cur_map)
n_common = len(common_idx)
# Discordant pairs for McNemar
# b: baseline correct, steered wrong ("regressions")
# c: baseline wrong, steered correct ("recoveries")
b = sum(1 for i in common_idx
if base_map[i] and not cur_map[i])
c = sum(1 for i in common_idx
if (not base_map[i]) and cur_map[i])
# McNemar p-value (exact binomial under H0: b ~ Bin(b+c, 0.5))
mcnemar_p = _mcnemar_pvalue(b, c)
base_acc = base["accuracy"]
cur_acc = cur["accuracy"]
delta = base_acc - cur_acc # positive => accuracy DROPPED
rel_drop = (delta / base_acc) if base_acc > 0 else 0.0
cur["vs_baseline"] = {
"baseline_accuracy": base_acc,
"steered_accuracy": cur_acc,
"absolute_drop": delta,
"relative_drop": rel_drop,
"n_common": n_common,
"n_regressions": b,
"n_recoveries": c,
"mcnemar_p_value": mcnemar_p,
"significant_at_0_05": (mcnemar_p is not None and mcnemar_p < 0.05),
}
log.info(
f" {config_name}/{ts_name}: "
f"acc {base_acc:.3f} -> {cur_acc:.3f} "
f"(Δ={delta:+.3f}, rel={rel_drop:+.1%}) "
f"regressions={b} recoveries={c} "
f"McNemar p={mcnemar_p if mcnemar_p is None else f'{mcnemar_p:.3g}'}"
)
write_json(results, DOWNSTREAM_ACC_JSON)
log.info(f"Saved: {DOWNSTREAM_ACC_JSON}")
if __name__ == "__main__":
main()
|