r1v.py

# Copyright 2024 Bytedance Ltd. and/or its affiliates
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import re
from typing import Dict

from mathruler.grader import grade_answer


def format_reward(predict: str) -> float:
    pattern = re.compile(r"<think>.*?</think>\s*<answer>.*?</answer>", re.DOTALL)
    format_match = re.fullmatch(pattern, predict)
    return 1.0 if format_match else 0.0


def accuracy_reward(predict: str, ground_truth: str) -> float:
    try:
        content_match = re.search(r"<answer>(.*?)</answer>", predict)
        given_answer = content_match.group(1).strip() if content_match else predict.strip()
        if grade_answer(given_answer, ground_truth.strip()):
            return 1.0

    except Exception:
        pass

    return 0.0


def compute_score(predict: str, ground_truth: str, format_weight: float = 0.5) -> Dict[str, float]:
    format_score = format_reward(predict)
    accuracy_score = accuracy_reward(predict, ground_truth)
    return {
        "overall": (1 - format_weight) * accuracy_score + format_weight * format_score,
        "format": format_score,
        "accuracy": accuracy_score,
    }