eqbench-ja/translate_eqbench.py

"""
translate_eqbench.py — EQ-Bench3 日本語化スクリプト

フロー:
    1. EQ-Bench3 の scenario_prompts.txt と scenario_notes.txt を読み込む
    2. vLLM（Qwen3.5-9B）経由で各シナリオを日本語に翻訳
    3. 翻訳結果を保存（チェックポイント対応・再開可能）
    4. 完成した日本語版ファイルを output/ に書き出し
    5. HF Hub にアップロード

実行例:
    python translate_eqbench.py
    python translate_eqbench.py --dry-run       # 最初の2シナリオのみ
    python translate_eqbench.py --target-file scenario_notes.txt  # ノートのみ
"""

from __future__ import annotations

import argparse
import asyncio
import json
import os
import sys
import time
import traceback
from datetime import datetime, timezone
from pathlib import Path

import httpx
from tqdm import tqdm

# ── 設定 ──────────────────────────────────────────────────────
VLLM_BASE_URL: str = os.environ.get("VLLM_BASE_URL", "http://localhost:8000/v1")
TRANSLATE_MODEL: str = os.environ.get("TRANSLATE_MODEL", "Qwen/Qwen3.5-9B")
HF_TOKEN: str = os.environ.get("HF_TOKEN", "")
HF_USERNAME: str = os.environ.get("HF_USERNAME", "YUGOROU")
HF_REPO: str = os.environ.get("HF_REPO", f"{HF_USERNAME}/teememo-eq-bench-ja")

EQBENCH_DIR: Path = Path(os.environ.get("EQBENCH_DIR", "/workspace/eqbench-ja/eqbench3"))
OUTPUT_DIR: Path = Path(os.environ.get("OUTPUT_DIR", "/workspace/eqbench-ja/output"))
CHECKPOINT_FILE: Path = Path(os.environ.get("CHECKPOINT_FILE", "/workspace/eqbench-ja/output/.checkpoint_translate.json"))

MAX_CONCURRENT: int = int(os.environ.get("MAX_CONCURRENT", "4"))
MAX_RETRIES: int = int(os.environ.get("MAX_RETRIES", "3"))
REQUEST_TIMEOUT: float = float(os.environ.get("REQUEST_TIMEOUT", "180.0"))

# ── 翻訳システムプロンプト ────────────────────────────────────
TRANSLATE_SYSTEM = """You are a professional translator specializing in Japanese localization of psychological and emotional intelligence assessment materials.

Your task is to translate English text into natural, fluent Japanese while:
1. Preserving the original meaning, nuance, and emotional tone precisely
2. Maintaining all formatting markers (######## , ####### , etc.) exactly as-is
3. Keeping scenario numbers, category names, and structural markers unchanged
4. Using natural conversational Japanese appropriate for the social context described
5. Preserving any special instructions in brackets [like this] translated into Japanese
6. For role-play scenarios involving interpersonal conflict, use natural Japanese speech patterns including appropriate keigo or casual speech as the context demands
7. Translating all proper nouns contextually (names can be kept or given Japanese equivalents)

Output ONLY the translated text with no explanations or commentary."""


# ── vLLM クライアント ─────────────────────────────────────────

class VLLMClient:
    def __init__(self) -> None:
        self._client = httpx.AsyncClient(
            timeout=httpx.Timeout(REQUEST_TIMEOUT),
        )
        self._semaphore = asyncio.Semaphore(MAX_CONCURRENT)

    async def close(self) -> None:
        await self._client.aclose()

    async def __aenter__(self) -> "VLLMClient":
        return self

    async def __aexit__(self, *_) -> None:
        await self.close()

    async def wait_for_server(self, max_wait: int = 300, interval: int = 5) -> None:
        print(f"[client] vLLM サーバーの起動を待機中 (最大 {max_wait}s)...")
        start = time.time()
        while time.time() - start < max_wait:
            try:
                response = await self._client.get(f"{VLLM_BASE_URL}/models")
                if response.status_code == 200:
                    print("[client] vLLM サーバーが起動しました。")
                    return
            except Exception:
                pass
            await asyncio.sleep(interval)
        raise TimeoutError(f"[client] vLLM サーバーが {max_wait}s 以内に起動しませんでした。")

    async def translate(self, text: str) -> str:
        """テキストを日本語に翻訳する。"""
        payload = {
            "model": TRANSLATE_MODEL,
            "messages": [
                {"role": "system", "content": TRANSLATE_SYSTEM},
                {"role": "user", "content": f"Translate the following to Japanese:\n\n{text}"},
            ],
            "temperature": 0.1,   # 翻訳は低温で安定させる
            "top_p": 0.9,
            "max_tokens": 4096,
            "chat_template_kwargs": {"enable_thinking": False},
        }

        last_exc = None
        for attempt in range(MAX_RETRIES):
            try:
                async with self._semaphore:
                    response = await self._client.post(
                        f"{VLLM_BASE_URL}/chat/completions",
                        json=payload,
                        headers={"Content-Type": "application/json"},
                    )
                response.raise_for_status()
                result = response.json()
                return result["choices"][0]["message"].get("content") or ""
            except Exception as exc:
                last_exc = exc
                wait = 2 ** attempt
                print(f"[client] リトライ {attempt+1}/{MAX_RETRIES} ({type(exc).__name__}) — {wait}s 待機")
                await asyncio.sleep(wait)
        raise RuntimeError(f"[client] 翻訳失敗: {last_exc}")


# ── scenario_prompts.txt パーサー ─────────────────────────────

def parse_scenario_prompts(filepath: Path) -> list[dict]:
    """
    scenario_prompts.txt を解析してシナリオリストを返す。

    Returns:
        [{"id": 1, "header": "...", "prompts": {"Prompt1": "...", "Prompt2": "..."}}, ...]
    """
    scenarios = []
    current: dict | None = None
    current_prompt_key: str | None = None
    current_prompt_lines: list[str] = []

    def _flush_prompt():
        if current and current_prompt_key and current_prompt_lines:
            current["prompts"][current_prompt_key] = "\n".join(current_prompt_lines).strip()
            current_prompt_lines.clear()

    with open(filepath, encoding="utf-8") as f:
        for line in f:
            line = line.rstrip("\n")

            # シナリオヘッダー: ######## 1 | Work Dilemma | ...
            if line.startswith("######## "):
                _flush_prompt()
                if current:
                    scenarios.append(current)
                parts = line.split("|")
                scenario_id = parts[0].replace("#", "").strip()
                current = {
                    "id": int(scenario_id) if scenario_id.isdigit() else scenario_id,
                    "header": line,
                    "prompts": {},
                }
                current_prompt_key = None
                current_prompt_lines = []

            # プロンプトキー: ####### Prompt1
            elif line.startswith("####### "):
                _flush_prompt()
                current_prompt_key = line.replace("#", "").strip()
                current_prompt_lines = []

            # プロンプト本文
            else:
                if current_prompt_key is not None:
                    current_prompt_lines.append(line)

    _flush_prompt()
    if current:
        scenarios.append(current)

    return scenarios


def parse_scenario_notes(filepath: Path) -> dict[str, str]:
    """
    scenario_notes.txt を解析する。
    # 1\n<note>\n# 2\n<note> の形式。

    Returns:
        {"1": "<note>", "2": "<note>", ...}
    """
    notes = {}
    current_key: str | None = None
    current_lines: list[str] = []

    def _flush():
        if current_key and current_lines:
            notes[current_key] = "\n".join(current_lines).strip()
            current_lines.clear()

    with open(filepath, encoding="utf-8") as f:
        for line in f:
            line = line.rstrip("\n")
            if line.startswith("# ") and line[2:].strip().isdigit():
                _flush()
                current_key = line[2:].strip()
                current_lines = []
            else:
                if current_key is not None:
                    current_lines.append(line)

    _flush()
    return notes


# ── チェックポイント ──────────────────────────────────────────

def load_checkpoint() -> dict:
    if CHECKPOINT_FILE.exists():
        try:
            return json.loads(CHECKPOINT_FILE.read_text(encoding="utf-8"))
        except Exception:
            pass
    return {"translated_scenarios": {}, "translated_notes": {}}


def save_checkpoint(state: dict) -> None:
    CHECKPOINT_FILE.parent.mkdir(parents=True, exist_ok=True)
    CHECKPOINT_FILE.write_text(json.dumps(state, ensure_ascii=False, indent=2), encoding="utf-8")


# ── 出力ファイル生成 ───────────────────────────────────────────

def build_output_prompts(
    scenarios: list[dict],
    translated: dict[str, dict],
) -> str:
    """翻訳済みシナリオを scenario_prompts.txt 形式に再構築する。"""
    lines = []
    for scenario in scenarios:
        sid = str(scenario["id"])
        if sid in translated:
            trans = translated[sid]
            lines.append(trans.get("header", scenario["header"]))
            for prompt_key, prompt_text in scenario["prompts"].items():
                lines.append(f"####### {prompt_key}")
                translated_text = trans.get("prompts", {}).get(prompt_key, prompt_text)
                lines.append(translated_text)
                lines.append("")
        else:
            # 未翻訳はオリジナルをそのまま
            lines.append(scenario["header"])
            for prompt_key, prompt_text in scenario["prompts"].items():
                lines.append(f"####### {prompt_key}")
                lines.append(prompt_text)
                lines.append("")
    return "\n".join(lines)


def build_output_notes(
    original_notes: dict[str, str],
    translated_notes: dict[str, str],
) -> str:
    """翻訳済みノートを scenario_notes.txt 形式に再構築する。"""
    lines = []
    for key, note in original_notes.items():
        lines.append(f"# {key}")
        lines.append(translated_notes.get(key, note))
        lines.append("")
    return "\n".join(lines)


# ── HF アップロード ───────────────────────────────────────────

async def upload_to_hf(output_dir: Path) -> None:
    if not HF_TOKEN:
        print("[hub] HF_TOKEN 未設定のためアップロードをスキップ")
        return
    try:
        from huggingface_hub import HfApi
        api = HfApi(token=HF_TOKEN)
        api.create_repo(repo_id=HF_REPO, repo_type="dataset", private=True, exist_ok=True)
        for f in output_dir.glob("*.txt"):
            api.upload_file(
                path_or_fileobj=str(f),
                path_in_repo=f"data/{f.name}",
                repo_id=HF_REPO,
                repo_type="dataset",
                commit_message=f"update: {f.name}",
            )
            print(f"[hub] アップロード完了: {f.name}")
        print(f"[hub] ✅ https://huggingface.co/datasets/{HF_REPO}")
    except Exception as e:
        print(f"[hub] アップロードエラー: {e}")
        print(traceback.format_exc())


# ── メイン処理 ────────────────────────────────────────────────

async def translate_scenarios(
    client: VLLMClient,
    scenarios: list[dict],
    state: dict,
    dry_run: bool,
    lock: asyncio.Lock,
) -> dict:
    """全シナリオを asyncio.gather で並列翻訳する。"""
    translated = state.get("translated_scenarios", {})
    target_scenarios = scenarios[:2] if dry_run else scenarios
    pending = [s for s in target_scenarios if str(s["id"]) not in translated]

    print(f"[translate] シナリオ翻訳開始: {len(target_scenarios)}件 (未翻訳: {len(pending)}件)")

    async def _translate_one(scenario: dict) -> None:
        sid = str(scenario["id"])
        trans_scenario: dict = {"header": scenario["header"], "prompts": {}}

        for prompt_key, prompt_text in scenario["prompts"].items():
            if not prompt_text.strip():
                trans_scenario["prompts"][prompt_key] = prompt_text
                continue
            try:
                translated_text = await client.translate(prompt_text)
                trans_scenario["prompts"][prompt_key] = translated_text
            except Exception as e:
                print(f"\n[translate] WARNING: シナリオ{sid}/{prompt_key} 翻訳失敗: {e}")
                trans_scenario["prompts"][prompt_key] = prompt_text

        async with lock:
            translated[sid] = trans_scenario
            state["translated_scenarios"] = translated
            save_checkpoint(state)
        print(f"[translate] シナリオ {sid} 完了")

    await asyncio.gather(*[_translate_one(s) for s in pending])

    print(f"[translate] シナリオ翻訳完了: {len(translated)}件")
    return translated


async def translate_notes(
    client: VLLMClient,
    notes: dict[str, str],
    state: dict,
    dry_run: bool,
    lock: asyncio.Lock,
) -> dict:
    """採点ノートを asyncio.gather で並列翻訳する。"""
    translated_notes = state.get("translated_notes", {})
    target_notes = dict(list(notes.items())[:2]) if dry_run else notes
    pending = {k: v for k, v in target_notes.items() if k not in translated_notes}

    print(f"[translate] ノート翻訳開始: {len(target_notes)}件 (未翻訳: {len(pending)}件)")

    async def _translate_one(key: str, note: str) -> None:
        try:
            translated_text = await client.translate(note)
        except Exception as e:
            print(f"\n[translate] WARNING: ノート#{key} 翻訳失敗: {e}")
            translated_text = note
        async with lock:
            translated_notes[key] = translated_text
            state["translated_notes"] = translated_notes
            save_checkpoint(state)
        print(f"[translate] ノート #{key} 完了")

    await asyncio.gather(*[_translate_one(k, v) for k, v in pending.items()])

    print(f"[translate] ノート翻訳完了: {len(translated_notes)}件")
    return translated_notes


async def main(dry_run: bool, target_file: str | None) -> None:
    start_time = datetime.now(timezone.utc)
    print(f"=== EQ-Bench3 日本語化開始 [{start_time.isoformat()}] ===")
    print(f"  EQ-Bench3 ディレクトリ: {EQBENCH_DIR}")
    print(f"  出力ディレクトリ       : {OUTPUT_DIR}")
    print(f"  翻訳モデル             : {TRANSLATE_MODEL}")
    print(f"  dry-run                : {dry_run}")
    print()

    OUTPUT_DIR.mkdir(parents=True, exist_ok=True)

    # ── データ読み込み ──────────────────────────────────────────
    prompts_file = EQBENCH_DIR / "data" / "scenario_prompts.txt"
    notes_file = EQBENCH_DIR / "data" / "scenario_notes.txt"

    if not prompts_file.exists():
        print(f"[ERROR] scenario_prompts.txt が見つかりません: {prompts_file}")
        print("        先に setup_translate.sh を実行してください。")
        sys.exit(1)

    print(f"[data] scenario_prompts.txt 読み込み中...")
    scenarios = parse_scenario_prompts(prompts_file)
    print(f"[data]   シナリオ数: {len(scenarios)}")

    print(f"[data] scenario_notes.txt 読み込み中...")
    notes = parse_scenario_notes(notes_file)
    print(f"[data]   ノート数: {len(notes)}")

    # ── チェックポイント読み込み ──────────────────────────────
    state = load_checkpoint()
    done_s = len(state.get("translated_scenarios", {}))
    done_n = len(state.get("translated_notes", {}))
    if done_s > 0 or done_n > 0:
        print(f"[checkpoint] 再開: シナリオ{done_s}件・ノート{done_n}件 処理済み")

    # ── vLLM 接続 ─────────────────────────────────────────────
    async with VLLMClient() as client:
        await client.wait_for_server()

        lock = asyncio.Lock()  # チェックポイント書き込みの競合を防ぐ

        # ── 翻訳実行 ────────────────────────────────────────────
        do_scenarios = target_file is None or target_file == "scenario_prompts.txt"
        do_notes = target_file is None or target_file == "scenario_notes.txt"

        if do_scenarios:
            translated_scenarios = await translate_scenarios(client, scenarios, state, dry_run, lock)
        else:
            translated_scenarios = state.get("translated_scenarios", {})

        if do_notes:
            translated_notes = await translate_notes(client, notes, state, dry_run, lock)
        else:
            translated_notes = state.get("translated_notes", {})

    # ── 出力ファイル生成 ─────────────────────────────────────
    print("[output] 出力ファイル生成中...")

    prompts_out = OUTPUT_DIR / "scenario_prompts_ja.txt"
    prompts_out.write_text(
        build_output_prompts(scenarios, translated_scenarios),
        encoding="utf-8",
    )
    print(f"[output]   ✅ {prompts_out}")

    notes_out = OUTPUT_DIR / "scenario_notes_ja.txt"
    notes_out.write_text(
        build_output_notes(notes, translated_notes),
        encoding="utf-8",
    )
    print(f"[output]   ✅ {notes_out}")

    # オリジナルもコピーしておく（比較用）
    import shutil
    shutil.copy(prompts_file, OUTPUT_DIR / "scenario_prompts_en.txt")
    shutil.copy(notes_file, OUTPUT_DIR / "scenario_notes_en.txt")

    # ── HF アップロード ───────────────────────────────────────
    if not dry_run:
        await upload_to_hf(OUTPUT_DIR)

    # チェックポイント削除
    if CHECKPOINT_FILE.exists() and not dry_run:
        CHECKPOINT_FILE.unlink()

    elapsed = (datetime.now(timezone.utc) - start_time).total_seconds()
    print()
    print(f"=== 翻訳完了 (所要時間: {elapsed/60:.1f}分) ===")
    print(f"  出力: {OUTPUT_DIR}/scenario_prompts_ja.txt")
    print(f"  出力: {OUTPUT_DIR}/scenario_notes_ja.txt")


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="EQ-Bench3 日本語化スクリプト")
    parser.add_argument("--dry-run", action="store_true", help="最初の2シナリオのみ処理して動作確認")
    parser.add_argument(
        "--target-file",
        choices=["scenario_prompts.txt", "scenario_notes.txt"],
        default=None,
        help="翻訳対象ファイルを指定（デフォルト: 両方）",
    )
    args = parser.parse_args()

    if not os.environ.get("HF_TOKEN"):
        print("[WARNING] HF_TOKEN が未設定です。HF アップロードはスキップされます。")

    asyncio.run(main(dry_run=args.dry_run, target_file=args.target_file))