reform_dataset_sqa.py

import os
import json
import time
import random
from typing import Dict, Any, List, Optional, Tuple

from openai import OpenAI
import multiprocessing as mp

# ============================================================
# IO helpers
# ============================================================
def load_json(file_path: str):
    with open(file_path, "r", encoding="utf-8") as f:
        return json.load(f)

def append_jsonl(path: str, obj: Dict[str, Any]):
    os.makedirs(os.path.dirname(path), exist_ok=True)
    with open(path, "a", encoding="utf-8") as f:
        f.write(json.dumps(obj, ensure_ascii=False) + "\n")

def load_done_keys(jsonl_path: str):
    done = set()
    if not os.path.exists(jsonl_path):
        return done
    with open(jsonl_path, "r", encoding="utf-8") as f:
        for line in f:
            line = line.strip()
            if not line:
                continue
            try:
                obj = json.loads(line)
                k = obj.get("key")
                if k is not None:
                    done.add(k)
            except Exception:
                continue
    return done

# ============================================================
# LLM helpers
# ============================================================
def safe_json_parse(text: str) -> Optional[Dict[str, Any]]:
    text = (text or "").strip()
    if text.startswith("```"):
        text = text.strip("`").strip()
        if text.lower().startswith("json"):
            text = text[4:].strip()
    try:
        return json.loads(text)
    except Exception:
        return None

def call_chat_json(
    client: OpenAI,
    model: str,
    system: str,
    user: str,
    max_tokens: int,
    temperature: float,
    top_p: float = 0.9,
    max_retries: int = 4,
) -> Dict[str, Any]:
    last_err = None
    for attempt in range(1, max_retries + 1):
        try:
            resp = client.chat.completions.create(
                model=model,
                messages=[
                    {"role": "system", "content": system},
                    {"role": "user", "content": user},
                ],
                temperature=temperature,
                top_p=top_p,
                max_tokens=max_tokens,
            )
            text = resp.choices[0].message.content
            obj = safe_json_parse(text)
            if obj is None:
                raise ValueError(f"JSON parse failed. Raw: {text[:200]}...")
            return obj
        except Exception as e:
            last_err = str(e)
            time.sleep(1.0 * attempt)
    raise RuntimeError(f"call_chat_json failed after {max_retries} retries. Last error: {last_err}")

def normalize_text(s: str) -> str:
    return " ".join((s or "").lower().split())


# ============================================================
# Binary question detection
# ============================================================
_BINARY_ANS = {"yes", "no", "true", "false"}
_AUX_START = {
    "is", "are", "was", "were",
    "do", "does", "did",
    "can", "could",
    "will", "would", "should",
    "has", "have", "had",
    "may", "might", "must",
}

def is_binary_qa(question: str, gt_answer: str) -> bool:
    a = normalize_text(gt_answer)
    if a in _BINARY_ANS:
        return True
    q = (question or "").strip().lower()
    # heuristic: starts with aux and looks like a yes/no question
    if q.endswith("?"):
        first = q.split()[0] if q.split() else ""
        if first in _AUX_START:
            return True
    return False

# ============================================================
# Round 1: situation + QA → correct statement
# ============================================================
SYSTEM_STATEMENT = r"""
You generate ONE correct answer sentence given a situation and a QA pair.

CRITICAL:
- The output must be a single, continuous sentence that integrates the situation text with a short answer clause that resolves the question (e.g., “When …, …”).

Return STRICT JSON: {"statement": "..."} only.

Rules:
- Must be accurate given the answer.
- Natural, concise, factual.
- Preserve referenced entity labels if any (e.g., "chair A").
- JSON only.
"""

USER_STATEMENT = """Situation: {situation}
Question: {question}
Ground-truth answer: {answer}

Return JSON only.
"""

def gen_statement(client: OpenAI, model: str, situation: str, question: str, answer: str) -> str:
    obj = call_chat_json(
        client=client,
        model=model,
        system=SYSTEM_STATEMENT,
        user=USER_STATEMENT.format(situation=situation, question=question, answer=answer),
        max_tokens=512,
        temperature=0.2,
    )
    st = obj.get("statement", "")
    if not isinstance(st, str) or not st.strip():
        raise ValueError(f"Bad statement: {obj}")
    return st.strip()

# ============================================================
# Round 2: wrong statements (N=9 normally, N=1 for binary)
# ============================================================
SYSTEM_WRONG_9 = r"""
You generate distractors for a ground-truth statement.

CRITICAL:
- EVERY sentence MUST integrates the situation text with a wrong answer clause (e.g., “When …, …”).

OUTPUT (STRICT JSON ONLY):
{{"distractors": ["<s1>", "<s2>", "<s3>", "<s4>", "<s5>", "<s6>", "<s7>", "<s8>", "<s9>"]}}

Rules:
- The situation should be fixed in each sentence.
- Exactly 9 strings, all unique.
- Each is a single declarative sentence.
- Each must be WRONG given the ground-truth answer.
- Preserve the same main subject/entity (same instance label if present).
- The distractors should be diverse, non-ambiguous, realistic. The difference between ground-truth and distractors should be VERY clear.
- No markdown, no extra keys.
"""

USER_WRONG_9 = """Situation: {situation}
Question: {question}
Ground-truth answer: {answer}
Correct statement: {correct_statement}

TASK
Generate 9 WRONG but plausible alternative declarative statements.

Return JSON only:
{{"distractors": ["...","...","...","...","...","...","...","...","..."]}}
"""

SYSTEM_WRONG_1 = r"""
You generate distractors for a ground-truth statement.

CRITICAL:
- EVERY sentence MUST integrates the situation text with a wrong answer clause (e.g., “When …, …”).

OUTPUT (STRICT JSON ONLY):
{{"distractors": ["<s1>"]}}

Rules:
- The situation should be fixed in the sentence.
- Exactly 1 string.
- The sentence must be WRONG given the ground-truth answer.
- Preserve the same main subject/entity (same instance label if present).
- Make the wrong answer VERY clear (not ambiguous).
- No markdown, no extra keys.
"""

USER_WRONG_1 = """Situation: {situation}
Question: {question}
Ground-truth answer: {answer}
Correct statement: {correct_statement}

TASK
Generate 1 WRONG but plausible alternative declarative statement.

Return JSON only:
{{"distractors": ["..."]}}
"""

def _add_unique(collected: List[str], seen: set, candidates: List[Any], situation: str) -> None:
    for x in candidates:
        if not isinstance(x, str):
            continue
        x = x.strip()
        if not x:
            continue
        x = x
        nx = normalize_text(x)
        if nx in seen:
            continue
        seen.add(nx)
        collected.append(x)

def gen_wrong_sentences(
    client: OpenAI,
    model: str,
    situation: str,
    question: str,
    answer: str,
    correct_statement: str,
    n: int = 9,
    batch_rounds: int = 5,
) -> List[str]:
    correct_norm = normalize_text(correct_statement)

    if n == 1:
        system = SYSTEM_WRONG_1
        user_tmpl = USER_WRONG_1
        max_tokens = 256
    else:
        system = SYSTEM_WRONG_9
        user_tmpl = USER_WRONG_9
        max_tokens = 1024

    for _ in range(batch_rounds):
        obj = call_chat_json(
            client=client,
            model=model,
            system=system,
            user=user_tmpl.format(
                situation=situation,
                question=question,
                answer=answer,
                correct_statement=correct_statement,
            ),
            max_tokens=max_tokens,
            temperature=0.95,
        )

        ds = obj.get("distractors", None)
        if not isinstance(ds, list):
            continue

        seen = {correct_norm}
        out: List[str] = []
        for s in ds:
            if not isinstance(s, str):
                continue
            s = s.strip()
            ns = normalize_text(s)
            if ns in seen:
                continue
            seen.add(ns)
            out.append(s)
            if len(out) >= n:
                return out[:n]

    raise ValueError(f"Could not collect {n} unique distractors in {batch_rounds} batch rounds.")

# ============================================================
# Build final options (1 correct + N wrong)
# ============================================================
def build_mcq(correct: str, wrongs: List[str], seed: Optional[int] = None):
    options = [correct] + list(wrongs)
    if seed is not None:
        rnd = random.Random(seed)
        rnd.shuffle(options)
    else:
        random.shuffle(options)
    label = options.index(correct)
    return options, label

# ============================================================
# Multiprocessing worker
# ============================================================
_WORKER_CLIENT = None
_WORKER_MODEL = None

def _init_worker(base_url: str, model_name: str, timeout: int = 3600):
    global _WORKER_CLIENT, _WORKER_MODEL
    _WORKER_MODEL = model_name
    _WORKER_CLIENT = OpenAI(api_key="EMPTY", base_url=base_url, timeout=timeout)

def _process_one(args: Tuple[str, int, str, str, str, str]) -> Dict[str, Any]:
    # key, idx, scene_id, situation, question, gt_answer
    key, idx, scene_id, situation, question, gt_answer = args
    try:
        correct_stmt = gen_statement(_WORKER_CLIENT, _WORKER_MODEL, situation, question, gt_answer)

        # binary => 1 wrong; otherwise 9 wrong
        n_wrong = 1 if is_binary_qa(question, gt_answer) else 9
        wrongs = gen_wrong_sentences(
            _WORKER_CLIENT,
            _WORKER_MODEL,
            situation,
            question,
            gt_answer,
            correct_stmt,
            n=n_wrong,
            batch_rounds=6 if n_wrong == 9 else 4,
        )

        seed = abs(hash(key)) % (2**32)
        options, label = build_mcq(correct_stmt, wrongs, seed=seed)

        out = {
            "key": key,
            "scene_id": scene_id,
            "situation": situation,
            "question": question,
            "gt_answer": gt_answer,
            "correct": correct_stmt,
            "options": options,   # 2 options for binary, 10 otherwise
            "label": label,
            "is_binary": (n_wrong == 1),
        }
        return {"ok": True, "out": out}

    except Exception as e:
        return {"ok": False, "err": {
            "key": key,
            "scene_id": scene_id,
            "situation": situation,
            "question": question,
            "gt_answer": gt_answer,
            "error": str(e),
        }}

# ============================================================
# Main (keeps your original data loading)
# ============================================================
def main():
    base_url = "http://lrc-alpha-sg-gpu06:22002/v1"
    model_name = "Qwen/Qwen3-VL-8B-Instruct"

    in_path = "/home/m50048399/transfered/ye_project/PointMapVerse/existing_datasets/ScanNet/annotations/sqa_task/balanced/v1_balanced_questions_test_scannetv2.json"
    ann_path = "/home/m50048399/transfered/ye_project/PointMapVerse/existing_datasets/ScanNet/annotations/sqa_task/balanced/v1_balanced_sqa_annotations_test_scannetv2.json"

    out_jsonl = "/home/m50048399/transfered/ye_project/PointMapVerse/derived/sqa_sentence_mcq_test_scannetv2.jsonl"
    err_jsonl = out_jsonl + ".errors.jsonl"

    qdata = load_json(in_path)["questions"]
    adata = load_json(ann_path)["annotations"]

    done = load_done_keys(out_jsonl)
    print(f"Loaded {len(qdata)} items. Already done: {len(done)}")

    tasks = []
    for idx, item in enumerate(qdata):
        scene_id = item.get("scene_id", "")
        question = (item.get("question") or "").strip()
        situation = (item.get("situation") or "").strip()

        gt_answer = ""
        try:
            gt_answer = (adata[idx]["answers"][0]["answer"] or "").strip()
        except Exception:
            gt_answer = ""

        if not situation or not question or not gt_answer:
            continue

        key = f"{scene_id}::{idx}"
        if key in done:
            continue

        tasks.append((key, idx, scene_id, situation, question, gt_answer))

    print(f"To process: {len(tasks)}")

    ctx = mp.get_context("spawn")
    with ctx.Pool(
        processes=8,
        initializer=_init_worker,
        initargs=(base_url, model_name, 3600),
        maxtasksperchild=50,
    ) as pool:
        processed = 0
        ok_cnt = 0
        err_cnt = 0

        for res in pool.imap_unordered(_process_one, tasks, chunksize=4):
            processed += 1
            if res["ok"]:
                append_jsonl(out_jsonl, res["out"])
                ok_cnt += 1
            else:
                append_jsonl(err_jsonl, res["err"])
                err_cnt += 1

            if processed % 100 == 0:
                print(f"Finished {processed}/{len(tasks)} | ok={ok_cnt} err={err_cnt}")

    print("Done.")

if __name__ == "__main__":
    main()