| """One official Avalon episode with its unmodified Session and game logic.""" |
| import asyncio |
| import json |
| import os |
| from pathlib import Path |
| import random |
| import time |
| from urllib.request import Request, urlopen |
|
|
| import numpy as np |
| from src.server.task import Session |
| from src.server.tasks.avalon import AvalonBench |
| from src.typings import AgentOutput |
| from response_format import SYSTEM_PROMPT, format_kind, validate_answer |
|
|
| OUT = Path(os.environ['EPISODE_OUT']) |
| OUT.mkdir(parents=True, exist_ok=False) |
| URL = os.environ['MODEL_URL'].rstrip('/') + '/v1/chat/completions' |
|
|
|
|
| async def main(): |
| random.seed(7) |
| np.random.seed(7) |
| task = AvalonBench(name='AvalonBench-dev-single', concurrency=1, num_players=5, |
| agent_list=['llm', 'naive', 'naive', 'naive', 'naive'], |
| discussion=False, data_file='data/avalon/dev.json', seed=7) |
| session = Session() |
| index = int(os.environ.get('EPISODE_INDEX', '0')) |
| game = asyncio.create_task(task.start_sample(index, session)) |
| response = None |
| calls = 0 |
| started = time.time() |
| result = {'task': 'avalon-dev-single', 'index': index, 'seed': 7, |
| 'agent_protocol': 'v3: format-only system prompt on structured extraction calls; repetition_penalty=1.1; strict validation; at most 3 attempts', |
| 'scope': 'in-process official game and agent session; separate GPU inference', |
| 'agentbench_commit': 'ed013ff9887b0c3d7864c56ae54d41eba54a99d8'} |
| try: |
| while not game.done(): |
| pull = asyncio.create_task(session.controller.agent_pull(response)) |
| done, _ = await asyncio.wait([pull, game], return_when=asyncio.FIRST_COMPLETED) |
| if game in done: |
| pull.cancel() |
| break |
| output = pull.result() |
| messages = [{'role': 'assistant' if h.role == 'agent' else h.role, |
| 'content': h.content} for h in output.history] |
| kind = format_kind(messages[-1]['content']) |
| if kind is not None: |
| messages.insert(0, {'role': 'system', 'content': |
| 'Format the requested answer from the provided text. Return only the exact ' |
| 'Answer: template with concrete values. No prose, markdown, or code fences.'}) |
| for attempt in range(3): |
| if calls >= 150: |
| raise RuntimeError('Model call limit reached') |
| payload = {'model': 'qwen3-coder-30b-awq', 'messages': messages, |
| 'temperature': 0, 'seed': 7, 'max_tokens': 768, |
| 'repetition_penalty': 1.1} |
| before = time.time() |
| with urlopen(Request(URL, data=json.dumps(payload).encode(), |
| headers={'Content-Type': 'application/json'}), timeout=180) as reply: |
| answer = json.load(reply) |
| calls += 1 |
| choice = answer['choices'][0] |
| content = choice['message']['content'] |
| error = None |
| try: |
| if choice['finish_reason'] != 'stop': |
| raise ValueError('Incomplete model response: ' + choice['finish_reason']) |
| validate_answer(kind, content) |
| except ValueError as exc: |
| error = str(exc) |
| with (OUT / 'model-calls.jsonl').open('a') as stream: |
| stream.write(json.dumps({'call': calls, 'started': before, |
| 'finished': time.time(), 'request': payload, |
| 'response': answer, 'format_kind': kind, |
| 'attempt': attempt + 1, 'validation_error': error}) + '\n') |
| if error is None: |
| response = AgentOutput(content=content) |
| break |
| if attempt == 2: |
| raise RuntimeError('Response format failed after 3 attempts: ' + error) |
| messages = messages + [ |
| {'role': 'assistant', 'content': content or ''}, |
| {'role': 'user', 'content': 'Correct only your response format. ' + error |
| + '. Preserve your chosen values. Return the requested answer and nothing else.'}] |
| result['episode'] = (await game).dict() |
| if result['episode']['status'] != 'completed': |
| raise RuntimeError('Benchmark did not complete: ' + str(result['episode']['status'])) |
| except BaseException as exc: |
| result['error'] = repr(exc) |
| raise |
| finally: |
| game.cancel() |
| result.update(model_calls=calls, elapsed_seconds=time.time()-started) |
| (OUT / 'result.json').write_text(json.dumps(result, indent=2, default=str)) |
| task.release() |
|
|
|
|
| asyncio.run(main()) |
|
|
