""" Inference Script — Meta_com OpenEnv Agent ========================================= MANDATORY - Before submitting, ensure the following variables are defined in your environment configuration: API_BASE_URL The API endpoint for the LLM. MODEL_NAME The model identifier to use for inference. HF_TOKEN Your Hugging Face / API key. LOCAL_IMAGE_NAME The name of the local image to use for the environment if you are using from_docker_image() method - Defaults are set only for API_BASE_URL and MODEL_NAME (and should reflect your active inference setup): API_BASE_URL = os.getenv("API_BASE_URL", "") MODEL_NAME = os.getenv("MODEL_NAME", "") - The inference script must be named `inference.py` and placed in the root directory of the project - Participants must use OpenAI Client for all LLM calls using above variables STDOUT FORMAT - The script must emit exactly three line types to stdout, in this order: [START] task= env= model= [STEP] step= action= reward=<0.00> done= error= [END] success= steps= score= rewards= TASK REQUIREMENTS - Must run at least 3 tasks with graders. - Each task score must be strictly between 0 and 1 (not 0.0 and not 1.0). """ import asyncio import os import textwrap from typing import Dict, List, Optional from openai import OpenAI from my_env_v4 import MyEnvV4Action, MyEnvV4Env # --------------------------------------------------------------------------- # Environment configuration (OpenEnv-compliant variable names) # --------------------------------------------------------------------------- API_BASE_URL = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1") MODEL_NAME = os.getenv("MODEL_NAME", "Qwen/Qwen2.5-72B-Instruct") HF_TOKEN = os.getenv("HF_TOKEN") # Optional — if you use from_docker_image(): LOCAL_IMAGE_NAME = os.getenv("LOCAL_IMAGE_NAME") BENCHMARK = os.getenv("MY_ENV_V4_BENCHMARK", "my_env_v4") # --------------------------------------------------------------------------- # Task registry — at least 3 tasks required by OpenEnv Phase 2 validation # --------------------------------------------------------------------------- TASKS: List[Dict] = [ { "id": "git_conflict_trivial", "max_steps": 4, "temperature": 0.5, "system_prompt": ( "You are resolving a trivial Git merge conflict. " "The conflict involves simple whitespace or formatting differences. " "Reply with exactly one corrected code block — no quotes, no prefixes." ), }, { "id": "git_conflict_multifile", "max_steps": 6, "temperature": 0.7, "system_prompt": ( "You are resolving a multi-file Git merge conflict. " "Two branches modified different parts of an API. Reconcile both changes. " "Reply with exactly one corrected code block — no quotes, no prefixes." ), }, { "id": "git_conflict_semantic", "max_steps": 8, "temperature": 0.8, "system_prompt": ( "You are resolving a deep semantic Git merge conflict. " "Two branches implement competing logic for the same feature. " "Synthesize both intents into a single coherent implementation. " "Reply with exactly one corrected code block — no quotes, no prefixes." ), }, ] MAX_TOKENS = 150 # Score boundary constants — OpenEnv requires scores strictly in (0, 1) SCORE_FLOOR = 0.001 SCORE_CEIL = 0.999 # --------------------------------------------------------------------------- # Structured logging helpers (stdout format required by OpenEnv) # --------------------------------------------------------------------------- def log_start(task: str, env: str, model: str) -> None: print(f"[START] task={task} env={env} model={model}", flush=True) def log_step(step: int, action: str, reward: float, done: bool, error: Optional[str]) -> None: error_val = error if error else "null" done_val = str(done).lower() print( f"[STEP] step={step} action={action} reward={reward:.2f} done={done_val} error={error_val}", flush=True, ) def log_end(success: bool, steps: int, score: float, rewards: List[float]) -> None: rewards_str = ",".join(f"{r:.2f}" for r in rewards) print( f"[END] success={str(success).lower()} steps={steps} score={score:.4f} rewards={rewards_str}", flush=True, ) # --------------------------------------------------------------------------- # Agent interaction helpers # --------------------------------------------------------------------------- def build_user_prompt(step: int, last_echoed: str, last_reward: float, history: List[str]) -> str: history_block = "\n".join(history[-4:]) if history else "None" return textwrap.dedent( f""" Step: {step} Last echoed message: {last_echoed!r} Last reward: {last_reward:.2f} Previous steps: {history_block} Send your next message. """ ).strip() def get_model_message( client: OpenAI, system_prompt: str, step: int, last_echoed: str, last_reward: float, history: List[str], temperature: float, ) -> str: user_prompt = build_user_prompt(step, last_echoed, last_reward, history) # Fallback heuristic when no real API key is available (build/test phase) if HF_TOKEN is None: return f"Conflict resolution patch for step {step}" try: completion = client.chat.completions.create( model=MODEL_NAME, messages=[ {"role": "system", "content": system_prompt}, {"role": "user", "content": user_prompt}, ], temperature=temperature, max_tokens=MAX_TOKENS, stream=False, ) text = (completion.choices[0].message.content or "").strip() return text if text else "hello" except Exception as exc: print(f"[DEBUG] Model request failed: {exc}", flush=True) return "hello" def clamp_score(raw: float) -> float: """Clamp a raw score to the open interval (0, 1) as required by OpenEnv.""" return min(max(raw, SCORE_FLOOR), SCORE_CEIL) # --------------------------------------------------------------------------- # Single-task episode runner # --------------------------------------------------------------------------- async def run_task(client: OpenAI, env: MyEnvV4Env, task: Dict) -> None: """Run one complete agent episode for the given task definition.""" task_id = task["id"] max_steps = task["max_steps"] temperature = task["temperature"] system_prompt = task["system_prompt"] # Max possible reward for normalization max_reward_per_step = MAX_TOKENS * 0.1 max_total_reward = max_steps * max_reward_per_step history: List[str] = [] rewards: List[float] = [] steps_taken = 0 score = 0.0 success = False log_start(task=task_id, env=BENCHMARK, model=MODEL_NAME) try: result = await env.reset(task_id=task_id) last_echoed = result.observation.echoed_message last_reward = 0.0 for step in range(1, max_steps + 1): if result.done: break message = get_model_message( client, system_prompt, step, last_echoed, last_reward, history, temperature ) result = await env.step(MyEnvV4Action(message=message)) obs = result.observation reward = result.reward or 0.0 done = result.done error = None rewards.append(reward) steps_taken = step last_echoed = obs.echoed_message last_reward = reward log_step(step=step, action=message, reward=reward, done=done, error=error) history.append(f"Step {step}: {message!r} -> reward {reward:+.2f}") if done: break raw_score = sum(rewards) / max_total_reward if max_total_reward > 0 else 0.5 score = clamp_score(raw_score) success = score >= 0.1 except Exception as exc: print(f"[DEBUG] Task {task_id} error: {exc}", flush=True) # Even on error, emit a valid clamped score so the grader accepts it score = clamp_score(0.0) success = False finally: log_end(success=success, steps=steps_taken, score=score, rewards=rewards) # --------------------------------------------------------------------------- # Main entrypoint — runs ALL registered tasks sequentially # --------------------------------------------------------------------------- async def main() -> None: api_key_to_use = HF_TOKEN if HF_TOKEN else "fake-key" client = OpenAI(base_url=API_BASE_URL, api_key=api_key_to_use) env = await MyEnvV4Env.from_docker_image(LOCAL_IMAGE_NAME) try: for task in TASKS: await run_task(client, env, task) finally: try: await env.close() except Exception as e: print(f"[DEBUG] env.close() error (container cleanup): {e}", flush=True) if __name__ == "__main__": asyncio.run(main())