(args)
| 437 | |
| 438 | |
| 439 | def main(args): |
| 440 | model, tokenizer = load_models_tokenizer(args) |
| 441 | |
| 442 | dev_result = {} |
| 443 | for subject_name in tqdm(TASK_NAME_MAPPING.keys()): |
| 444 | val_file_path = os.path.join( |
| 445 | args.eval_data_path, "val", f"{subject_name}_val.csv" |
| 446 | ) |
| 447 | dev_file_path = os.path.join( |
| 448 | args.eval_data_path, "dev", f"{subject_name}_dev.csv" |
| 449 | ) |
| 450 | # test_file_path = os.path.join(args.eval_data_path, 'test', f'{subject_name}_test.csv') |
| 451 | val_df = pd.read_csv(val_file_path) |
| 452 | dev_df = pd.read_csv(dev_file_path) |
| 453 | # test_df = pd.read_csv(test_file_path) |
| 454 | |
| 455 | score = eval_subject( |
| 456 | model, |
| 457 | tokenizer, |
| 458 | subject_name, |
| 459 | val_df, |
| 460 | dev_df=dev_df, |
| 461 | k=5, |
| 462 | few_shot=True, |
| 463 | save_result_dir=f"outs/ceval_eval_result", |
| 464 | ) |
| 465 | dev_result[subject_name] = score |
| 466 | cal_ceval(dev_result) |
| 467 | |
| 468 | |
| 469 | if __name__ == "__main__": |
no test coverage detected