example_submission.py

"""
WM Bench — Example Submission Script
=====================================
Any world model can participate in WM Bench using this template.
No 3D environment needed — text input/output only.

Usage:
  python example_submission.py --api_url YOUR_MODEL_API --api_key YOUR_KEY --model YOUR_MODEL_NAME
"""

import json
import argparse
import requests
import time
from pathlib import Path

# ── 설정 ────────────────────────────────────────────────────────
DATASET_PATH = Path(__file__).parent.parent / "data" / "wm_bench_dataset.json"

SYSTEM_PROMPT = """You are a world model. Given scene_context as JSON, respond in exactly 2 lines:
Line 1: PREDICT: left=<safe|danger>(<reason>), right=<safe|danger>(<reason>), fwd=<safe|danger>(<reason>), back=<safe|danger>(<reason>)
Line 2: MOTION: <describe the character's physical motion and emotional state in one sentence>
Respond ONLY these 2 lines. No explanation."""

# ── 메인 평가 함수 ───────────────────────────────────────────────
def run_evaluation(api_url: str, api_key: str, model: str, output_path: str = "my_submission.json"):
    """
    WM Bench 평가를 실행하고 제출 파일을 생성합니다.
    
    Parameters:
        api_url: OpenAI 호환 API URL (예: https://api.openai.com/v1/chat/completions)
        api_key: API 키
        model: 모델 이름
        output_path: 제출 파일 경로
    """
    # 데이터셋 로드
    with open(DATASET_PATH, "r", encoding="utf-8") as f:
        dataset = json.load(f)
    
    scenarios = dataset["scenarios"]
    print(f"✅ 데이터셋 로드: {len(scenarios)}개 시나리오")
    print(f"🤖 모델: {model}")
    print(f"🔗 API: {api_url}\n")

    results = []
    errors = 0

    for i, scenario in enumerate(scenarios):
        sc_id = scenario["id"]
        cat = scenario["cat"]
        scene = scenario["scene_context"]

        # API 호출
        t0 = time.time()
        response_text, latency_ms = call_api(api_url, api_key, model, scene)
        
        if response_text is None:
            errors += 1
            print(f"  ❌ {sc_id} ({cat}): API 오류")
            results.append({
                "id": sc_id,
                "cat": cat,
                "response": None,
                "latency_ms": latency_ms,
                "error": True
            })
        else:
            results.append({
                "id": sc_id,
                "cat": cat,
                "response": response_text,
                "latency_ms": round(latency_ms, 1),
                "error": False
            })
            
            if (i + 1) % 10 == 0:
                print(f"  ✓ {i+1}/100 완료 ({cat})")

    # 제출 파일 생성
    submission = {
        "model": model,
        "api_url": api_url,
        "track": "A",  # Text-Only Track
        "submitted_at": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()),
        "total_scenarios": len(scenarios),
        "errors": errors,
        "results": results
    }

    with open(output_path, "w", encoding="utf-8") as f:
        json.dump(submission, f, ensure_ascii=False, indent=2)

    print(f"\n✅ 제출 파일 생성: {output_path}")
    print(f"   총 시나리오: {len(scenarios)}, 오류: {errors}")
    print(f"\n📤 다음 단계: WM Bench Space에 제출 파일 업로드")
    print(f"   https://huggingface.co/spaces/FINAL-Bench/worldmodel-bench")
    return output_path


def call_api(api_url: str, api_key: str, model: str, scene_context: dict):
    """OpenAI 호환 API 호출"""
    headers = {
        "Content-Type": "application/json",
        "Authorization": f"Bearer {api_key}"
    }
    payload = {
        "model": model,
        "max_tokens": 200,
        "temperature": 0.0,
        "messages": [
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": f"scene_context: {json.dumps(scene_context)}"}
        ]
    }
    t0 = time.time()
    try:
        r = requests.post(api_url, headers=headers, json=payload, timeout=30)
        r.raise_for_status()
        text = r.json()["choices"][0]["message"]["content"]
        return text, (time.time() - t0) * 1000
    except Exception as e:
        return None, (time.time() - t0) * 1000


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="WM Bench Submission Script")
    parser.add_argument("--api_url", required=True, help="OpenAI-compatible API URL")
    parser.add_argument("--api_key", required=True, help="API Key")
    parser.add_argument("--model", required=True, help="Model name")
    parser.add_argument("--output", default="my_submission.json", help="Output file path")
    args = parser.parse_args()
    
    run_evaluation(args.api_url, args.api_key, args.model, args.output)