File size: 10,529 Bytes
f6c2fe1 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 | #!/usr/bin/env python3
import argparse
import json
import re
import time
from pathlib import Path
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers import StoppingCriteria, StoppingCriteriaList
BOXED_CHOICE_RE = re.compile(r"\\boxed\s*\{\s*([ABCD])\s*\}")
class StopAfterFirstBoxedChoice(StoppingCriteria):
def __init__(self, tokenizer, prompt_len):
self.tokenizer = tokenizer
self.prompt_len = prompt_len
def __call__(self, input_ids, scores, **kwargs):
gen_text = self.tokenizer.decode(input_ids[0][self.prompt_len:], skip_special_tokens=True)
return BOXED_CHOICE_RE.search(gen_text) is not None
class StepwiseState:
def __init__(self, tokenizer):
self.tokenizer = tokenizer
self.active = False
self.prefill_done = False
def update_from_input_ids(self, input_ids):
# During prefill, do not intervene.
if input_ids.shape[-1] > 1:
self.active = False
self.prefill_done = True
return
# During autoregressive decoding, intervene only when previous token ended a step.
tail = self.tokenizer.decode(input_ids[0][-8:], skip_special_tokens=False)
self.active = tail.endswith("\n\n") or tail.endswith("\n \n")
def read_jsonl(path):
rows = []
with open(path, "r", encoding="utf-8") as f:
for line in f:
if line.strip():
rows.append(json.loads(line))
return rows
def append_jsonl(path, obj):
with open(path, "a", encoding="utf-8") as f:
f.write(json.dumps(obj, ensure_ascii=False) + "\n")
def build_prompt(tokenizer, problem):
messages = [{"role": "user", "content": problem}]
if hasattr(tokenizer, "apply_chat_template") and tokenizer.chat_template is not None:
return tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
return problem
def extract_first_boxed_choice(text):
m = BOXED_CHOICE_RE.search(text)
return m.group(1) if m else None
def truncate_after_first_boxed_choice(text):
m = BOXED_CHOICE_RE.search(text)
return text[:m.end()] if m else text
def think_tokens_like_crest(tokenizer, cot):
seg = cot.split("</think>", 1)[0] if "</think>" in cot else cot
return len(tokenizer.encode(seg, add_special_tokens=False))
def repetition_score(text, n=40):
if not text:
return 0.0
chunks = [text[i:i+n] for i in range(0, max(0, len(text) - n + 1), n)]
if not chunks:
return 0.0
counts = {}
for c in chunks:
counts[c] = counts.get(c, 0) + 1
return max(counts.values()) / max(1, len(chunks))
def mon_total(text):
markers = [
"wait", "Wait", "check", "Check", "rethink", "Rethink",
"however", "However", "maybe", "Maybe", "alternatively",
"Actually", "actually", "Let's verify", "verify",
]
return sum(text.count(m) for m in markers)
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--data", required=True)
ap.add_argument("--model-path", required=True)
ap.add_argument("--direction-path", required=True)
ap.add_argument("--out", required=True)
ap.add_argument("--seed", type=int, default=64)
ap.add_argument("--lambda-weight", type=float, default=-0.48)
ap.add_argument("--layers", default="6-57")
ap.add_argument("--max-new-tokens", type=int, default=32768)
ap.add_argument("--temperature", type=float, default=0.6)
ap.add_argument("--top-p", type=float, default=0.95)
ap.add_argument("--limit", type=int, default=None)
ap.add_argument("--force", action="store_true")
ap.add_argument("--stop-after-boxed", action="store_true")
args = ap.parse_args()
torch.manual_seed(args.seed)
out_path = Path(args.out)
out_path.parent.mkdir(parents=True, exist_ok=True)
if args.force and out_path.exists():
out_path.unlink()
rows = read_jsonl(args.data)
if args.limit is not None:
rows = rows[:args.limit]
existing = set()
if out_path.exists():
for line in out_path.read_text(encoding="utf-8").splitlines():
if line.strip():
existing.add(json.loads(line).get("_key"))
l0, l1 = [int(x) for x in args.layers.split("-")]
layers = list(range(l0, l1 + 1))
print(f"[ReflCtrl] model_path = {args.model_path}", flush=True)
print(f"[ReflCtrl] data = {args.data}", flush=True)
print(f"[ReflCtrl] direction = {args.direction_path}", flush=True)
print(f"[ReflCtrl] out = {args.out}", flush=True)
print(f"[ReflCtrl] n = {len(rows)}", flush=True)
print(f"[ReflCtrl] seed = {args.seed}", flush=True)
print(f"[ReflCtrl] lambda = {args.lambda_weight}", flush=True)
print(f"[ReflCtrl] layers = {args.layers}", flush=True)
print(f"[ReflCtrl] max_new_tokens = {args.max_new_tokens}", flush=True)
print(f"[ReflCtrl] temperature = {args.temperature}", flush=True)
print(f"[ReflCtrl] top_p = {args.top_p}", flush=True)
print(f"[ReflCtrl] stop_after_boxed = {args.stop_after_boxed}", flush=True)
print(f"[ReflCtrl] existing records = {len(existing)}", flush=True)
tokenizer = AutoTokenizer.from_pretrained(args.model_path, trust_remote_code=True, use_fast=True)
if tokenizer.pad_token_id is None:
tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForCausalLM.from_pretrained(
args.model_path,
trust_remote_code=True,
torch_dtype="auto",
device_map="auto",
)
model.eval()
ckpt = torch.load(args.direction_path, map_location="cpu")
components = ckpt["components"]
state = StepwiseState(tokenizer)
original_forward = model.forward
def wrapped_forward(*f_args, **f_kwargs):
input_ids = f_kwargs.get("input_ids", None)
if input_ids is None and len(f_args) > 0:
input_ids = f_args[0]
if input_ids is not None:
state.update_from_input_ids(input_ids)
return original_forward(*f_args, **f_kwargs)
model.forward = wrapped_forward
handles = []
def make_hook(name):
vec = components[name]["mean_diff"].float()
def hook(module, inputs, output):
if not state.active:
return output
h = output[0] if isinstance(output, tuple) else output
v = vec.to(device=h.device, dtype=h.dtype)
h2 = h.clone()
h2[:, -1, :] = h2[:, -1, :] + args.lambda_weight * v
if isinstance(output, tuple):
return (h2,) + output[1:]
return h2
return hook
for l in layers:
for suffix, module in [
("self_attn", model.model.layers[l].self_attn),
("mlp", model.model.layers[l].mlp),
]:
name = f"model.layers[{l}].{suffix}"
if name in components:
handles.append(module.register_forward_hook(make_hook(name)))
else:
print(f"[WARN] missing direction component: {name}", flush=True)
first_device = next(model.parameters()).device
try:
for i, item in enumerate(rows):
key = f"P{i}_ReflCtrl_lam{args.lambda_weight:+.2f}".replace("+", "p").replace("-", "m").replace(".", "p")
if key in existing:
print(f"[SKIP] {key}", flush=True)
continue
problem = item.get("problem") or item.get("question") or item.get("prompt")
if problem is None:
raise KeyError(f"missing problem/question/prompt at row {i}")
gt = str(item.get("answer") or item.get("gt") or item.get("label")).strip().upper()
if gt and gt[0] in "ABCD":
gt = gt[0]
prompt = build_prompt(tokenizer, problem)
inputs = tokenizer(prompt, return_tensors="pt")
inputs = {k: v.to(first_device) for k, v in inputs.items()}
input_len = inputs["input_ids"].shape[-1]
stopping = None
if args.stop_after_boxed:
stopping = StoppingCriteriaList([StopAfterFirstBoxedChoice(tokenizer, input_len)])
state.active = False
state.prefill_done = False
t0 = time.time()
with torch.no_grad():
out = model.generate(
**inputs,
max_new_tokens=args.max_new_tokens,
do_sample=True,
temperature=args.temperature,
top_p=args.top_p,
pad_token_id=tokenizer.pad_token_id,
eos_token_id=tokenizer.eos_token_id,
stopping_criteria=stopping,
)
gen_ids = out[0][input_len:]
cot = tokenizer.decode(gen_ids, skip_special_tokens=True)
if args.stop_after_boxed:
cot = truncate_after_first_boxed_choice(cot)
pred = extract_first_boxed_choice(cot)
correct = pred == gt if pred is not None and gt else False
ttok = think_tokens_like_crest(tokenizer, cot)
rep = repetition_score(cot)
mtot = mon_total(cot)
elapsed = time.time() - t0
collapse = pred is None or rep >= 0.50 or ttok >= args.max_new_tokens - 8
rec = {
"_key": key,
"method": "ReflCtrl",
"problem_idx": i,
"lambda": args.lambda_weight,
"layers": args.layers,
"seed": args.seed,
"problem": problem,
"cot": cot,
"pred": pred,
"gt": gt,
"correct": correct,
"has_boxed": pred is not None,
"think_tokens": ttok,
"n_chars": len(cot),
"mon_total": mtot,
"repetition_score": rep,
"collapse": collapse,
"elapsed_s": elapsed,
}
append_jsonl(out_path, rec)
status = "OK" if correct else "BAD"
print(
f"{key} pred={pred} gt={gt} {status} "
f"tok={ttok} mon={mtot} rep={rep:.2f} "
f"collapse={collapse} t={elapsed:.1f}s",
flush=True,
)
finally:
for h in handles:
h.remove()
if __name__ == "__main__":
main()
|