(pred_str, data_name, use_last_number=True)
| 348 | |
| 349 | |
| 350 | def extract_answer(pred_str, data_name, use_last_number=True): |
| 351 | pred_str = pred_str.replace("\u043a\u0438", "") |
| 352 | |
| 353 | if "final answer is $" in pred_str and "$. I hope" in pred_str: |
| 354 | # minerva_math |
| 355 | tmp = pred_str.split("final answer is $", 1)[1] |
| 356 | pred = tmp.split("$. I hope", 1)[0].strip() |
| 357 | elif "boxed" in pred_str: |
| 358 | ans = pred_str.split("boxed")[-1] |
| 359 | if len(ans) == 0: |
| 360 | return "" |
| 361 | elif ans[0] == "{": |
| 362 | stack = 1 |
| 363 | a = "" |
| 364 | for c in ans[1:]: |
| 365 | if c == "{": |
| 366 | stack += 1 |
| 367 | a += c |
| 368 | elif c == "}": |
| 369 | stack -= 1 |
| 370 | if stack == 0: |
| 371 | break |
| 372 | a += c |
| 373 | else: |
| 374 | a += c |
| 375 | else: |
| 376 | a = ans.split("$")[0].strip() |
| 377 | pred = a |
| 378 | elif "he answer is" in pred_str: |
| 379 | pred = pred_str.split("he answer is")[-1].strip() |
| 380 | elif "final answer is" in pred_str: |
| 381 | pred = pred_str.split("final answer is")[-1].strip() |
| 382 | elif "答案是" in pred_str: |
| 383 | # Handle Chinese few-shot multiple choice problem answer extraction |
| 384 | pred = pred_str.split("答案是")[1].strip().split("\n\n")[0].strip() |
| 385 | else: # use the last number |
| 386 | if use_last_number: |
| 387 | pattern = "-?\d*\.?\d+" |
| 388 | pred = re.findall(pattern, pred_str.replace(",", "")) |
| 389 | if len(pred) >= 1: |
| 390 | pred = pred[-1] |
| 391 | else: |
| 392 | pred = "" |
| 393 | else: |
| 394 | pred = "" |
| 395 | |
| 396 | # choice answer |
| 397 | if ( |
| 398 | data_name in ["sat_math", "aqua"] |
| 399 | or "mmlu" in data_name |
| 400 | ): |
| 401 | tmp = re.findall(r"\b(A|B|C|D|E)\b", pred.upper()) |
| 402 | if tmp: |
| 403 | pred = tmp[-1] |
| 404 | else: |
| 405 | pred = pred.strip().strip(".") |
| 406 | |
| 407 | # multiple line |
no test coverage detected