| import hashlib |
| import math |
| import json |
| from dataclasses import asdict |
| from public_context import feedback_text |
| def choose(predictions): |
| if len(predictions) != 5: |
| raise ValueError('Expected four candidates and one current program') |
| utility = [row['utility'] for row in predictions] |
| if not all((math.isfinite(v) and 0 <= v <= 1.25 for v in utility)): |
| raise ValueError('Invalid utility') |
| best = max(range(4), key=lambda i: utility[i]) |
| return best if utility[best] > utility[4] else 4 |
|
|
| class Predictor: |
|
|
| def __init__(self, checkpoint, backbone): |
| from semantic_world import SemanticExecutionWorld |
| self.model = SemanticExecutionWorld.load(checkpoint, backbone, device='cuda').eval() |
| self.identity = self.model.identity |
|
|
| def predict(self, problem, code, feedback, tests, candidates): |
| public = [asdict(test) for test in tests if test.test_id.startswith('public_')] |
| observed = asdict(feedback) |
| observed['results'] = [row for row in observed['results'] if row['test_id'].startswith('public_')] |
| contexts = [f'Whole-training-suite execution forecast. Only public observations are supplied.\nProblem:\n{problem}\nCurrent program:\n{code}\nTrusted previous public execution:\n{feedback_text(observed)}\nCandidate action:\n{candidate.code}\nOfficial public test suite:\n' + json.dumps(public, ensure_ascii=False) for candidate in candidates] |
| values = self.model.score_suite(contexts) |
| return [dict(candidate_id=c.candidate_id, utility=float(values['utility'][i]), full_suite_pass_probability=float(values['full_suite_pass_probability'][i]), predicted_pass_fraction=float(values['predicted_pass_fraction'][i])) for (i, c) in enumerate(candidates)] |
|
|
| def state_key(code, feedback, step): |
| outcomes = [{k: getattr(row, k) for k in ('test_id', 'status', 'actual', 'expected', 'stderr')} for row in feedback.results] |
| return hashlib.sha256(json.dumps(dict(code=code, feedback=outcomes, step=step), sort_keys=True).encode()).hexdigest() |
|
|
