vedkdev/FlakyTestSleuthOpenEnvRL

1"""2Inference Script Example3===================================4MANDATORY5- Before submitting, ensure the following variables are defined in your environment configuration:6    API_BASE_URL   The API endpoint for the LLM.7    MODEL_NAME     The model identifier to use for inference.8    HF_TOKEN       Your Hugging Face / API key.9    LOCAL_IMAGE_NAME The name of the local image to use for the environment if you are using from_docker_image()10                     method11 12- Defaults are set only for API_BASE_URL and MODEL_NAME 13    (and should reflect your active inference setup):14    API_BASE_URL = os.getenv("API_BASE_URL", "<your-active-endpoint>")15    MODEL_NAME = os.getenv("MODEL_NAME", "<your-active-model>")16    17- The inference script must be named `inference.py` and placed in the root directory of the project18- Participants must use OpenAI Client for all LLM calls using above variables19 20STDOUT FORMAT21- The script must emit exactly three line types to stdout, in this order:22 23    [START] task=<task_name> env=<benchmark> model=<model_name>24    [STEP]  step=<n> action=<action_str> reward=<0.00> done=<true|false> error=<msg|null>25    [END]   success=<true|false> steps=<n> score=<score> rewards=<r1,r2,...,rn>26 27  Rules:28    - One [START] line at episode begin.29    - One [STEP] line per step, immediately after env.step() returns.30    - One [END] line after env.close(), always emitted (even on exception).31    - reward and rewards are formatted to 2 decimal places.32    - done and success are lowercase booleans: true or false.33    - error is the raw last_action_error string, or null if none.34    - All fields on a single line with no newlines within a line.35    - Each tasks should return score in [0, 1]36 37  Example:38    [START] task=click-test env=miniwob model=Qwen3-VL-30B39    [STEP] step=1 action=click('123') reward=0.00 done=false error=null40    [STEP] step=2 action=fill('456','text') reward=0.00 done=false error=null41    [STEP] step=3 action=click('789') reward=1.00 done=true error=null42    [END] success=true steps=3 score=1.00 rewards=0.00,0.00,1.0043"""44 45import asyncio46import os47import textwrap48from typing import List, Optional49 50from openai import OpenAI51 52from my_env_v4 import MyEnvV4Action, MyEnvV4Env53IMAGE_NAME = os.getenv("IMAGE_NAME") # If you are using docker image 54API_KEY = os.getenv("HF_TOKEN") or os.getenv("API_KEY")55 56API_BASE_URL = os.getenv("API_BASE_URL") or "https://router.huggingface.co/v1"57MODEL_NAME = os.getenv("MODEL_NAME") or "Qwen/Qwen2.5-72B-Instruct"58TASK_NAME = os.getenv("MY_ENV_V4_TASK", "echo")59BENCHMARK = os.getenv("MY_ENV_V4_BENCHMARK", "my_env_v4")60MAX_STEPS = 861TEMPERATURE = 0.762MAX_TOKENS = 15063SUCCESS_SCORE_THRESHOLD = 0.1  # normalized score in [0, 1]64 65# Max possible reward: each token contributes 0.1, across all steps66_MAX_REWARD_PER_STEP = MAX_TOKENS * 0.167MAX_TOTAL_REWARD = MAX_STEPS * _MAX_REWARD_PER_STEP68 69SYSTEM_PROMPT = textwrap.dedent(70    """71    You are interacting with a simple echo environment.72    Each turn you must send a message. The environment will echo it back.73    Reward is proportional to message length: reward = len(message) * 0.174    Your goal is to maximize total reward by sending meaningful, substantive messages.75    Reply with exactly one message string — no quotes, no prefixes, just the message text.76    """77).strip()78 79 80def log_start(task: str, env: str, model: str) -> None:81    print(f"[START] task={task} env={env} model={model}", flush=True)82 83 84def log_step(step: int, action: str, reward: float, done: bool, error: Optional[str]) -> None:85    error_val = error if error else "null"86    done_val = str(done).lower()87    print(88        f"[STEP] step={step} action={action} reward={reward:.2f} done={done_val} error={error_val}",89        flush=True,90    )91 92 93def log_end(success: bool, steps: int, score: float, rewards: List[float]) -> None:94    rewards_str = ",".join(f"{r:.2f}" for r in rewards)95    print(f"[END] success={str(success).lower()} steps={steps} score={score:.3f} rewards={rewards_str}", flush=True)96 97 98def build_user_prompt(step: int, last_echoed: str, last_reward: float, history: List[str]) -> str:99    history_block = "\n".join(history[-4:]) if history else "None"100    return textwrap.dedent(101        f"""102        Step: {step}103        Last echoed message: {last_echoed!r}104        Last reward: {last_reward:.2f}105        Previous steps:106        {history_block}107        Send your next message.108        """109    ).strip()110 111 112def get_model_message(client: OpenAI, step: int, last_echoed: str, last_reward: float, history: List[str]) -> str:113    user_prompt = build_user_prompt(step, last_echoed, last_reward, history)114    try:115        completion = client.chat.completions.create(116            model=MODEL_NAME,117            messages=[118                {"role": "system", "content": SYSTEM_PROMPT},119                {"role": "user", "content": user_prompt},120            ],121            temperature=TEMPERATURE,122            max_tokens=MAX_TOKENS,123            stream=False,124        )125        text = (completion.choices[0].message.content or "").strip()126        return text if text else "hello"127    except Exception as exc:128        print(f"[DEBUG] Model request failed: {exc}", flush=True)129        return "hello"130 131 132async def main() -> None:133    client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)134 135    env = await MyEnvV4Env.from_docker_image(IMAGE_NAME)136 137    history: List[str] = []138    rewards: List[float] = []139    steps_taken = 0140    score = 0.0141    success = False142 143    log_start(task=TASK_NAME, env=BENCHMARK, model=MODEL_NAME)144 145    try:146        result = await env.reset() # OpenENV.reset()147        last_echoed = result.observation.echoed_message148        last_reward = 0.0149 150        for step in range(1, MAX_STEPS + 1):151            if result.done:152                break153 154            message = get_model_message(client, step, last_echoed, last_reward, history)155 156            result = await env.step(MyEnvV4Action(message=message))157            obs = result.observation158 159            reward = result.reward or 0.0160            done = result.done161            error = None162 163            rewards.append(reward)164            steps_taken = step165            last_echoed = obs.echoed_message166            last_reward = reward167 168            log_step(step=step, action=message, reward=reward, done=done, error=error)169 170            history.append(f"Step {step}: {message!r} -> reward {reward:+.2f}")171 172            if done:173                break174 175        score = sum(rewards) / MAX_TOTAL_REWARD if MAX_TOTAL_REWARD > 0 else 0.0176        score = min(max(score, 0.001), 0.999)  # clamp to (0.001, 0.999)177        success = score >= SUCCESS_SCORE_THRESHOLD178 179    finally:180        try:181            await env.close()182        except Exception as e:183            print(f"[DEBUG] env.close() error (container cleanup): {e}", flush=True)184        log_end(success=success, steps=steps_taken, score=score, rewards=rewards)185 186 187if __name__ == "__main__":188    asyncio.run(main())