Team Ai
Apppublic

Neethish05/meta-pytorch-hackathon

sourceHugging Faceupdated 6mo agoView on Hugging Face
0likes
validate.py57 linesDownload Raw Back to root
1"""2Validation script for AI Code Review Arena (OpenEnv stack).3"""4 5import asyncio6 7from openenv_environment import OpenEnvCodeReviewArena8from openenv_models import ActionType, CodeReviewAction9from tasks import get_all_tasks10 11 12async def test_basic_functionality() -> None:13    env = OpenEnvCodeReviewArena()14    state = await env.reset(task_id=0)15    assert state.observation.step_count == 016    assert state.observation.files17    action = CodeReviewAction(type=ActionType.COMMENT, line=3, message="Index out of bounds error")18    state = await env.step(action)19    assert state.reward is not None20    assert isinstance(state.reward.value, float)21 22 23async def test_all_tasks() -> None:24    tasks = get_all_tasks()25    assert len(tasks) >= 326    for idx, _ in enumerate(tasks[:3]):27        env = OpenEnvCodeReviewArena()28        state = await env.reset(task_id=idx)29        assert state.observation.files30        assert state.observation.current_issues31 32 33async def test_grader_range() -> None:34    env = OpenEnvCodeReviewArena()35    await env.reset(task_id=2)36    state = await env.step(CodeReviewAction(type=ActionType.COMMENT, line=2, message="security injection bug"))37    _ = state38    score = env.get_task_grade()39    assert 0.0 <= score <= 1.040 41 42async def run_validation() -> bool:43    try:44        await test_basic_functionality()45        await test_all_tasks()46        await test_grader_range()47        print("Validation passed.")48        return True49    except Exception as exc:50        print(f"Validation failed: {exc}")51        return False52 53 54if __name__ == "__main__":55    ok = asyncio.run(run_validation())56    raise SystemExit(0 if ok else 1)57