Download scripts/eval_answer_utils.py from oof-baroomf/research-logs: direct link, hf CLI and curl.
- Browser
- Download file 3.05 kB
-
https://huggingface.co/datasets/oof-baroomf/research-logs/resolve/main/scripts/eval_answer_utils.py
- Command line
-
hf download hf://datasets/oof-baroomf/research-logs/scripts/eval_answer_utils.py
-
curl -L -o eval_answer_utils.py https://huggingface.co/datasets/oof-baroomf/research-logs/resolve/main/scripts/eval_answer_utils.py
3.05 kB
| from __future__ import annotations | |
| import re | |
| FINAL_LINE_RE = re.compile(r"(?mi)^\s*FINAL:\s*(.+?)\s*$") | |
| FINAL_MARKER_RE = re.compile(r"(?i)FINAL:\s*([^\n<]+)") | |
| FINAL_ANSWER_LINE_RE = re.compile( | |
| r"(?mi)^\s*(?:\**\s*)?(?:final answer|the final answer|answer)\s*(?:is|=|:)\s*([^\n]+)$" | |
| ) | |
| BOXED_RE = re.compile(r"\\boxed\{([^{}]+)\}", flags=re.IGNORECASE) | |
| NUMERIC_RE = re.compile(r"-?\d+(?:\.\d+)?") | |
| def normalize(text: str) -> str: | |
| text = text.strip() | |
| text = re.sub(r"\s+", " ", text) | |
| return text.casefold() | |
| def clean_answer_text(text: str) -> str: | |
| text = re.sub(r"</?answer>", " ", text, flags=re.IGNORECASE) | |
| text = re.sub(r"\\boxed\{([^{}]*)\}", r"\1", text) | |
| text = text.replace("$", " ") | |
| text = re.sub(r"[*`_#]+", " ", text) | |
| text = re.sub(r"^\**\s*final answer\s*(?:is|:)?\s*", "", text, flags=re.IGNORECASE) | |
| text = re.sub(r"^\**\s*the final answer\s*(?:is|:)?\s*", "", text, flags=re.IGNORECASE) | |
| text = re.sub(r"^\**\s*final\s*:\s*", "", text, flags=re.IGNORECASE) | |
| text = re.sub(r"^\**\s*answer\s*(?:is|:)?\s*", "", text, flags=re.IGNORECASE) | |
| text = re.sub(r"\s+", " ", text) | |
| return text.strip(" .:,;!?)(") | |
| def extract_final_answer(text: str) -> str: | |
| for regex in [FINAL_LINE_RE, FINAL_MARKER_RE, BOXED_RE, FINAL_ANSWER_LINE_RE]: | |
| matches = regex.findall(text) | |
| for match in reversed(matches): | |
| cleaned = clean_answer_text(match) | |
| if cleaned: | |
| return cleaned | |
| lines = [line.strip() for line in text.strip().splitlines() if line.strip()] | |
| for line in reversed(lines): | |
| cleaned = clean_answer_text(line.removeprefix("FINAL:").strip()) | |
| if cleaned: | |
| return cleaned | |
| return clean_answer_text(text) | |
| def has_final_marker(text: str) -> bool: | |
| return ( | |
| FINAL_LINE_RE.search(text) is not None | |
| or FINAL_MARKER_RE.search(text) is not None | |
| or FINAL_ANSWER_LINE_RE.search(text) is not None | |
| ) | |
| def normalize_numeric_string(text: str) -> str: | |
| value = text.strip() | |
| negative = value.startswith("-") | |
| if negative: | |
| value = value[1:] | |
| if "." in value: | |
| whole, fractional = value.split(".", 1) | |
| whole = whole.lstrip("0") or "0" | |
| fractional = fractional.rstrip("0") | |
| if not fractional: | |
| normalized = whole | |
| else: | |
| normalized = f"{whole}.{fractional}" | |
| else: | |
| normalized = value.lstrip("0") or "0" | |
| if normalized == "0": | |
| negative = False | |
| if negative: | |
| return f"-{normalized}" | |
| return normalized | |
| def answers_match(predicted: str, gold: str) -> bool: | |
| if normalize(predicted) == normalize(gold): | |
| return True | |
| gold_is_number = NUMERIC_RE.fullmatch(gold.strip()) is not None | |
| if gold_is_number: | |
| normalized_gold = normalize_numeric_string(gold) | |
| predicted_numbers = NUMERIC_RE.findall(predicted) | |
| if predicted_numbers and normalize_numeric_string(predicted_numbers[-1]) == normalized_gold: | |
| return True | |
| return False | |