scripts/00_generate_contrastive_cots.py

"""
Stage 00 (v8b): Generate 200 CONTRASTIVE pairs of CoTs with Qwen3-8B.

For each math problem, generate two CoTs from the same base model using
two different system prompts:

  HIGH-REFLECTION: encourage verification, second-guessing, strategy
                   switching, look-back.
  LOW-REFLECTION:  discourage second-guessing; commit to the first
                   approach and only verify at the very end.

Both CoTs go to RAW_COTS_PATH as JSONL with fields:
  problem, high_reflection_cot, low_reflection_cot, high_full, low_full

Resume: skip problems already in RAW_COTS_PATH; append new ones.
"""
import argparse, json, os, sys, time, random
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))

import torch
from tqdm import tqdm
from configs.paths import MATH_SOURCE_PATH, RAW_COTS_PATH, LOG_DIR, ensure_dirs
from src.utils import (
    build_chat_prompt, get_device, load_model_and_tokenizer,
    read_jsonl, setup_logger,
)

N_SAMPLE = 200
MAX_TOKENS = 8192
SEED = 42

SYSTEM_HIGH = (
    "You are a careful math problem solver. Think step by step. "
    "After each major step, briefly pause and ASK YOURSELF whether the "
    "step is correct. If you are unsure, reconsider, try a different "
    "approach, and verify your reasoning before continuing. Cross-check "
    "intermediate results. After you reach an answer, look back over "
    "your whole solution and confirm it makes sense."
)
SYSTEM_LOW = (
    "You are a confident math problem solver. Think step by step. "
    "Commit to the first reasonable approach you see and follow it "
    "through without second-guessing. Do not revisit earlier steps. "
    "Do not consider alternative methods. State your final answer "
    "directly at the end."
)


def get_problem(item):
    for k in ("problem", "question", "query", "input"):
        if k in item and item[k]:
            return item[k]
    return ""


def _gen(model, tokenizer, system: str, problem: str, device: str, max_new: int):
    prompt = build_chat_prompt(tokenizer, problem,
                               enable_thinking=True, system=system)
    inp = tokenizer(prompt, return_tensors="pt",
                    truncation=True, max_length=2048).to(device)
    with torch.no_grad():
        out = model.generate(
            **inp, max_new_tokens=max_new, do_sample=False,
            temperature=1.0, top_p=1.0,
            pad_token_id=tokenizer.eos_token_id,
        )
    full = tokenizer.decode(out[0], skip_special_tokens=True)
    prompt_text = tokenizer.decode(inp["input_ids"][0], skip_special_tokens=True)
    if full.startswith(prompt_text):
        return full[len(prompt_text):]
    return full


def _extract_thinking(text: str) -> str:
    if "</think>" in text:
        text = text[:text.index("</think>")]
    if text.strip().startswith("<think>"):
        text = text.strip()[len("<think>"):]
    return text.strip()


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--n-sample", type=int, default=N_SAMPLE)
    ap.add_argument("--max-tokens", type=int, default=MAX_TOKENS)
    ap.add_argument("--seed", type=int, default=SEED)
    ap.add_argument("--force", action="store_true")
    args = ap.parse_args()

    ensure_dirs("monitoring")
    log = setup_logger("00_gen_contrastive",
                       os.path.join(LOG_DIR, "00_gen_contrastive.log"))
    log.info("=" * 70)
    log.info(f"Stage 00 (v8b contrastive): generate {args.n_sample} CoT PAIRS")
    log.info(f"  MATH_SOURCE_PATH = {MATH_SOURCE_PATH}")
    log.info(f"  RAW_COTS_PATH    = {RAW_COTS_PATH}")
    log.info("=" * 70)

    if not os.path.exists(MATH_SOURCE_PATH):
        log.error(f"source problems not found: {MATH_SOURCE_PATH}"); sys.exit(1)

    all_items = read_jsonl(MATH_SOURCE_PATH)
    problems = [get_problem(it) for it in all_items if get_problem(it)]
    random.seed(args.seed)
    sampled = random.sample(problems, min(args.n_sample, len(problems)))
    log.info(f"  sampled {len(sampled)} problems")

    existing = {}
    if os.path.exists(RAW_COTS_PATH) and not args.force:
        for rec in read_jsonl(RAW_COTS_PATH):
            p = rec.get("problem")
            if p and rec.get("high_reflection_cot") and rec.get("low_reflection_cot"):
                existing[p] = rec
        log.info(f"  [resume] {len(existing)} pairs already on disk")

    todo = [p for p in sampled if p not in existing]
    log.info(f"  to generate: {len(todo)}")
    if not todo:
        log.info("All pairs already generated — DONE.")
        return

    device = get_device()
    log.info("Loading model...")
    model, tokenizer = load_model_and_tokenizer(device=device)

    out_fh = open(RAW_COTS_PATH, "a", encoding="utf-8")
    for i, prob in enumerate(tqdm(todo, desc=" generate pairs")):
        t0 = time.time()
        try:
            full_hi = _gen(model, tokenizer, SYSTEM_HIGH, prob, device, args.max_tokens)
            full_lo = _gen(model, tokenizer, SYSTEM_LOW, prob, device, args.max_tokens)
            hi = _extract_thinking(full_hi)
            lo = _extract_thinking(full_lo)
            rec = {
                "problem": prob,
                "high_reflection_cot": hi,
                "low_reflection_cot": lo,
                "high_full": full_hi,
                "low_full": full_lo,
                "gen_time_s": time.time() - t0,
            }
            out_fh.write(json.dumps(rec, ensure_ascii=False) + "\n")
            out_fh.flush()
            log.info(f"  [{i+1}/{len(todo)}] hi_len={len(hi)} lo_len={len(lo)} "
                     f"t={time.time()-t0:.0f}s")
        except Exception as e:
            log.warning(f"  [{i+1}] failed: {e}")
            continue

    out_fh.close()
    log.info(f"Saved (appended) -> {RAW_COTS_PATH}")
    log.info("Done.")


if __name__ == "__main__":
    main()