Blame for code/e5cb5ca/src/classify_qwen3guard.py

"""Qwen3Guard-Gen による有害性分類（監査の第2手法）。

辞書・ブロックリスト・共起規則とは独立した分類器として、目視サンプルや
境界文書に適用する。全件へ適用するには計算量が大きいため、
``sample_for_review.py`` が出力するサンプルJSONLを入力にする想定。

入力JSONL: ``{"id": ..., "url": ..., "text": ..., "match_positions": [..]}``
出力JSONL: 入力に ``safety``（Safe/Unsafe/Controversial）、``categories``、
``raw_output``、``elapsed_s`` を加えたもの。併せて ``<output>.manifest.json`` に
モデル名・リビジョン・実行条件を記録する。

依存関係は ``uv sync --group classifier`` で導入する（torch, transformers）。

    uv run python abeja-cc-ja/src/classify_qwen3guard.py --input samples.jsonl --output classified.jsonl
"""

from __future__ import annotations

import argparse
import json
import re
import time
from datetime import datetime, timezone
from pathlib import Path
from typing import Any


DEFAULT_MODEL = "Qwen/Qwen3Guard-Gen-0.6B"
SAFETY_RE = re.compile(r"Safety:\s*(Safe|Unsafe|Controversial)")
CATEGORY_RE = re.compile(
    r"(Violent|Non-violent Illegal Acts|Sexual Content or Sexual Acts|PII|Suicide & Self-Harm|"
    r"Unethical Acts|Politically Sensitive Topics|Copyright Violation|Jailbreak|None)"
)
# Qwen3Guard のカテゴリと監査項目（5〜10）の対応。
CATEGORY_TO_AUDIT_ITEM = {
    "Sexual Content or Sexual Acts": 5,
    "Violent": 6,
    "Suicide & Self-Harm": 6,
    "Unethical Acts": 7,
    "Non-violent Illegal Acts": 8,
    "Politically Sensitive Topics": 9,
    "Copyright Violation": 10,
}


def build_excerpt(text: str, positions: list[int] | None, max_chars: int, window: int = 300, head: int = 1000) -> str:
    """先頭と、一致位置の前後を優先して抜粋する。長文をそのまま渡さない。"""

    if len(text) <= max_chars:
        return text
    pieces: list[tuple[int, int]] = [(0, min(head, len(text)))]
    for position in (positions or [])[:4]:
        pieces.append((max(0, position - window), min(len(text), position + window)))
    pieces.sort()
    merged: list[list[int]] = []
    for start, end in pieces:
        if merged and start <= merged[-1][1]:
            merged[-1][1] = max(merged[-1][1], end)
        else:
            merged.append([start, end])
    excerpt = " … ".join(text[start:end] for start, end in merged)
    return excerpt[:max_chars]


def parse_output(content: str) -> tuple[str | None, list[str]]:
    safety = SAFETY_RE.search(content)
    categories = [item for item in CATEGORY_RE.findall(content) if item != "None"]
    return (safety.group(1) if safety else None), sorted(set(categories))


def load_model(model_name: str, device: str):
    import torch
    from transformers import AutoModelForCausalLM, AutoTokenizer

    tokenizer = AutoTokenizer.from_pretrained(model_name)
    kwargs: dict[str, Any] = {}
    if device == "auto":
        kwargs["device_map"] = "auto"
    try:
        model = AutoModelForCausalLM.from_pretrained(model_name, dtype="auto", **kwargs)
    except TypeError:
        model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", **kwargs)
    if device != "auto":
        model = model.to(device)
    model.eval()
    return tokenizer, model, torch


def classify_text(tokenizer, model, torch, text: str, max_new_tokens: int = 64) -> str:
    messages = [{"role": "user", "content": text}]
    prompt = tokenizer.apply_chat_template(messages, tokenize=False)
    inputs = tokenizer([prompt], return_tensors="pt").to(model.device)
    with torch.no_grad():
        generated = model.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
    output_ids = generated[0][len(inputs.input_ids[0]):].tolist()
    return tokenizer.decode(output_ids, skip_special_tokens=True)


def main(argv: list[str] | None = None) -> None:
    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    parser.add_argument("--input", type=Path, required=True)
    parser.add_argument("--output", type=Path, required=True)
    parser.add_argument("--model", default=DEFAULT_MODEL)
    parser.add_argument("--device", default="auto", help="auto / cpu / mps / cuda")
    parser.add_argument("--max-chars", type=int, default=3000)
    parser.add_argument("--max-new-tokens", type=int, default=64)
    parser.add_argument("--limit", type=int, default=None)
    args = parser.parse_args(argv)

    tokenizer, model, torch = load_model(args.model, args.device)
    revision = getattr(getattr(model, "config", None), "_commit_hash", None)
    started = datetime.now(timezone.utc).isoformat()
    counts: dict[str, int] = {"Safe": 0, "Unsafe": 0, "Controversial": 0, "unparsed": 0}
    processed = 0
    args.output.parent.mkdir(parents=True, exist_ok=True)
    with args.input.open(encoding="utf-8") as reader, args.output.open("w", encoding="utf-8") as writer:
        for line in reader:
            if not line.strip():
                continue
            if args.limit is not None and processed >= args.limit:
                break
            record = json.loads(line)
            text = build_excerpt(record.get("text") or "", record.get("match_positions"), args.max_chars)
            t0 = time.perf_counter()
            raw = classify_text(tokenizer, model, torch, text, args.max_new_tokens)
            elapsed = time.perf_counter() - t0
            safety, categories = parse_output(raw)
            counts[safety or "unparsed"] += 1
            record.update({
                "safety": safety,
                "categories": categories,
                "audit_items": sorted({CATEGORY_TO_AUDIT_ITEM[c] for c in categories if c in CATEGORY_TO_AUDIT_ITEM}),
                "raw_output": raw.strip(),
                "excerpt_chars": len(text),
                "elapsed_s": round(elapsed, 3),
                "model": args.model,
            })
            record.pop("text", None)
            writer.write(json.dumps(record, ensure_ascii=False, separators=(",", ":")) + "\n")
            processed += 1
    manifest = {
        "model": args.model,
        "model_revision": revision,
        "device": str(model.device),
        "max_chars": args.max_chars,
        "max_new_tokens": args.max_new_tokens,
        "started_at_utc": started,
        "finished_at_utc": datetime.now(timezone.utc).isoformat(),
        "input": str(args.input),
        "output": str(args.output),
        "processed": processed,
        "counts": counts,
        "category_to_audit_item": CATEGORY_TO_AUDIT_ITEM,
        "taxonomy_note": "Qwen3Guard: Safe / Controversial / Unsafe。Controversial は文脈依存の判定。",
    }
    Path(str(args.output) + ".manifest.json").write_text(json.dumps(manifest, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    print(json.dumps({"processed": processed, "counts": counts, "model": args.model}, ensure_ascii=False))


if __name__ == "__main__":
    main()