| 80 | return aggregate_results |
| 81 | |
| 82 | def parse_args(): |
| 83 | parser = argparse.ArgumentParser(description="Evaluate theorem proof verification results") |
| 84 | |
| 85 | # File paths |
| 86 | parser.add_argument("--input_file", default="/workspace/ky_ding/math/verify/0411/verified_stp_3200.json", |
| 87 | help="Path to verification results file") |
| 88 | parser.add_argument("--output_file", default="/workspace/ky_ding/math/verify/0411/verified_stp_3200_success_rate_0414.json", |
| 89 | help="Path to evaluation results file") |
| 90 | |
| 91 | # Evaluation parameters |
| 92 | parser.add_argument("--n_simulations", default=50, type=int, |
| 93 | help="Number of Monte Carlo simulations") |
| 94 | parser.add_argument("--n_processes", default=50, type=int, |
| 95 | help="Number of processes for Monte Carlo simulation") |
| 96 | parser.add_argument("--custom_sample_sizes", default=None, type=str, |
| 97 | help="Comma-separated list of custom sample sizes (e.g., '1,5,10,50,100')") |
| 98 | |
| 99 | return parser.parse_args() |
| 100 | |
| 101 | def main(): |
| 102 | args = parse_args() |