reflctrl/scripts/run_reflctrl_math500_single.py

#!/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


CREST_SYSTEM = (
    "Answer the following questions. You should think step-by-step "
    "and put your final answer within \\boxed{}."
)


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": "system", "content": CREST_SYSTEM},
        {"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 CREST_SYSTEM + "\n\n" + problem


def first_complete_boxed_end(text):
    i = text.find("\\boxed")
    if i < 0:
        return None
    j = text.find("{", i)
    if j < 0:
        return None
    depth = 0
    for k in range(j, len(text)):
        if text[k] == "{":
            depth += 1
        elif text[k] == "}":
            depth -= 1
            if depth == 0:
                return k + 1
    return None


def truncate_after_first_boxed(text):
    end = first_complete_boxed_end(text)
    return text[:end] if end is not None else text


def last_boxed(text):
    if not text:
        return None
    i = text.rfind("\\boxed")
    if i < 0:
        return None
    j = text.find("{", i)
    if j < 0:
        return None
    depth = 0
    for k in range(j, len(text)):
        if text[k] == "{":
            depth += 1
        elif text[k] == "}":
            depth -= 1
            if depth == 0:
                return text[j + 1:k].strip()
    return None


class StopAfterFirstBoxedMath(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 first_complete_boxed_end(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):
        if input_ids.shape[-1] > 1:
            self.active = False
            self.prefill_done = True
            return
        tail = self.tokenizer.decode(input_ids[0][-8:], skip_special_tokens=False)
        self.active = tail.endswith("\n\n") or tail.endswith("\n \n")


def _norm(s):
    if s is None:
        return ""
    t = str(s).strip()
    for x in ["\\left", "\\right", "\\!", "\\,", "\\;", "$", " "]:
        t = t.replace(x, "")
    return t.lower()


def _as_int(s):
    if s is None:
        return None
    t = re.sub(r"[^\d\-]", "", str(s))
    try:
        return int(t)
    except Exception:
        return None


def _as_float(s):
    if s is None:
        return None
    try:
        return float(str(s).replace(",", "").replace("$", ""))
    except Exception:
        return None


def _latex_to_sympy_src(s):
    t = str(s)
    t = t.replace("\\dfrac", "\\frac")
    t = re.sub(r"\\frac\{([^{}]+)\}\{([^{}]+)\}", r"((\1)/(\2))", t)
    t = re.sub(r"\\sqrt\{([^{}]+)\}", r"sqrt(\1)", t)
    t = re.sub(r"\\sqrt\s*(\d+)", r"sqrt(\1)", t)
    t = t.replace("\\cdot", "*").replace("\\times", "*")
    t = t.replace("^", "**")
    t = re.sub(r"\\pi\b", "pi", t)
    t = re.sub(r"\\(left|right|!|,|;|:)", "", t)
    t = re.sub(r"\\[a-zA-Z]+", "", t)
    t = t.replace("{", "(").replace("}", ")").replace("$", "")
    return t


def _sympy_eq(a, b):
    try:
        from sympy import sympify, simplify
    except ImportError:
        return None
    try:
        pa = sympify(_latex_to_sympy_src(a))
        pb = sympify(_latex_to_sympy_src(b))
        return bool(simplify(pa - pb) == 0)
    except Exception:
        return None


def is_correct(pred, gold):
    if pred is None or gold is None or not str(gold).strip():
        return False
    p, g = str(pred).strip(), str(gold).strip()
    if p == g:
        return True
    pi, gi = _as_int(p), _as_int(g)
    if pi is not None and gi is not None and "/" not in p and "/" not in g:
        return pi == gi
    pf, gf = _as_float(p), _as_float(g)
    if pf is not None and gf is not None and abs(pf) < 1e9 and abs(gf) < 1e9:
        if abs(pf - gf) < 1e-6:
            return True
    sym = _sympy_eq(p, g)
    if sym is not None:
        return sym
    np_, ng_ = _norm(p), _norm(g)
    return np_ == ng_ and np_ != ""


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=0)
    ap.add_argument("--lambda-weight", type=float, default=-0.48)
    ap.add_argument("--layers", default="6-47")
    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)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(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-MATH500] model_path       = {args.model_path}", flush=True)
    print(f"[ReflCtrl-MATH500] data             = {args.data}", flush=True)
    print(f"[ReflCtrl-MATH500] direction        = {args.direction_path}", flush=True)
    print(f"[ReflCtrl-MATH500] out              = {args.out}", flush=True)
    print(f"[ReflCtrl-MATH500] n                = {len(rows)}", flush=True)
    print(f"[ReflCtrl-MATH500] seed             = {args.seed}", flush=True)
    print(f"[ReflCtrl-MATH500] lambda           = {args.lambda_weight}", flush=True)
    print(f"[ReflCtrl-MATH500] layers           = {args.layers}", flush=True)
    print(f"[ReflCtrl-MATH500] max_new_tokens   = {args.max_new_tokens}", flush=True)
    print(f"[ReflCtrl-MATH500] temperature      = {args.temperature}", flush=True)
    print(f"[ReflCtrl-MATH500] top_p            = {args.top_p}", flush=True)
    print(f"[ReflCtrl-MATH500] stop_after_boxed = {args.stop_after_boxed}", flush=True)
    print(f"[ReflCtrl-MATH500] 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") or "").strip()

            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([StopAfterFirstBoxedMath(tokenizer, input_len)])

            state.active = False
            state.prefill_done = False

            gen_seed = args.seed * 1000 + i
            torch.manual_seed(gen_seed)
            if torch.cuda.is_available():
                torch.cuda.manual_seed_all(gen_seed)

            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(cot)

            pred = last_boxed(cot)
            correct = is_correct(pred, gt)
            ttok = think_tokens_like_crest(tokenizer, cot)
            rep = repetition_score(cot)
            mtot = mon_total(cot)
            elapsed = time.time() - t0
            has_boxed = pred is not None
            collapse = (not has_boxed) or rep >= 0.5 or ttok >= args.max_new_tokens

            rec = {
                "_key": key,
                "problem_idx": i,
                "seed": args.seed,
                "lambda_weight": args.lambda_weight,
                "layers": args.layers,
                "problem": problem,
                "cot": cot,
                "pred": pred,
                "gt": gt,
                "correct": correct,
                "has_boxed": has_boxed,
                "think_tokens": ttok,
                "n_chars": len(cot),
                "mon_total": mtot,
                "repetition_score": rep,
                "collapse": collapse,
                "elapsed_s": elapsed,
            }
            append_jsonl(out_path, rec)

            print(
                f"P{i}_ReflCtrl_lam{args.lambda_weight:+.2f} "
                f"pred={pred} gt={gt} {'OK' if correct else 'BAD'} "
                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()