(example: Dict[str, Any], data_name)
| 534 | |
| 535 | |
| 536 | def parse_ground_truth(example: Dict[str, Any], data_name): |
| 537 | if "gt_cot" in example and "gt" in example: |
| 538 | if data_name in ["math"]: |
| 539 | gt_ans = extract_answer(example["gt_cot"], data_name) |
| 540 | elif data_name in STRIP_EXCEPTIONS: |
| 541 | gt_ans = example["gt"] |
| 542 | else: |
| 543 | gt_ans = strip_string(example["gt"]) |
| 544 | return example["gt_cot"], gt_ans |
| 545 | |
| 546 | # parse ground truth |
| 547 | if data_name in ["math", "minerva_math", "math500"]: #关键代码 |
| 548 | gt_cot = example["solution"] |
| 549 | gt_ans = extract_answer(gt_cot, data_name) |
| 550 | elif data_name == "gsm8k": |
| 551 | gt_cot, gt_ans = example["answer"].split("####") |
| 552 | elif data_name == "svamp": |
| 553 | gt_cot, gt_ans = example["Equation"], example["Answer"] |
| 554 | elif data_name == "asdiv": |
| 555 | gt_cot = example["formula"] |
| 556 | gt_ans = re.sub(r"\(.*?\)", "", example["answer"]) |
| 557 | elif data_name == "mawps": |
| 558 | gt_cot, gt_ans = None, example["target"] |
| 559 | elif data_name == "tabmwp": |
| 560 | gt_cot = example["solution"] |
| 561 | gt_ans = example["answer"] |
| 562 | if example["ans_type"] in ["integer_number", "decimal_number"]: |
| 563 | if "/" in gt_ans: |
| 564 | gt_ans = int(gt_ans.split("/")[0]) / int(gt_ans.split("/")[1]) |
| 565 | elif "," in gt_ans: |
| 566 | gt_ans = float(gt_ans.replace(",", "")) |
| 567 | elif "%" in gt_ans: |
| 568 | gt_ans = float(gt_ans.split("%")[0]) / 100 |
| 569 | else: |
| 570 | gt_ans = float(gt_ans) |
| 571 | elif data_name == "carp_en": |
| 572 | gt_cot, gt_ans = example["steps"], example["answer"] |
| 573 | elif data_name == "mmlu_stem": |
| 574 | abcd = "ABCD" |
| 575 | gt_cot, gt_ans = None, abcd[example["answer"]] |
| 576 | elif data_name == "sat_math": |
| 577 | gt_cot, gt_ans = None, example["Answer"] |
| 578 | elif data_name == "aqua": |
| 579 | gt_cot, gt_ans = None, example["correct"] |
| 580 | elif data_name in ["gaokao2023en", "college_math", "gaokao_math_cloze"]: |
| 581 | gt_cot, gt_ans = None, example["answer"].replace("$", "").strip() |
| 582 | elif data_name == "gaokao_math_qa": |
| 583 | gt_cot, gt_ans = None, example["label"] |
| 584 | elif data_name in ["gaokao2024_mix", "cn_middle_school"]: |
| 585 | if len(example["choice_answer"]) > 0: |
| 586 | gt_cot, gt_ans = None, example["choice_answer"] |
| 587 | else: |
| 588 | gt_cot, gt_ans = None, example["answer"] |
| 589 | elif data_name == "olympiadbench": |
| 590 | gt_cot, gt_ans = None, example["final_answer"][0].strip("$") |
| 591 | elif data_name in [ |
| 592 | "aime24", |
| 593 | "aime25", |
no test coverage detected