| """ |
| Number Game Reward Function |
| |
| 评分规则： |
| - 选择正确的数字: +1.0 |
| - 选择错误的数字: 0.0 |
| |
| 输入格式: |
| reward_input = { |
| "response": "1", # 模型输出的答案 (0/1/2) |
| "response_length": 10, # 响应长度（token数） |
| "ground_truth": "1" # 正确答案 (0/1/2) |
| } |
| |
| 输出格式: |
| { |
| "overall": 1.0, # 总分（必需字段） |
| "accuracy": 1.0 # 准确率（可选，用于监控） |
| } |
| """ |
|
|
| import re |
| from typing import Any |
|
|
|
|
| |
| REWARD_NAME = "number_game" |
| REWARD_TYPE = "batch" |
|
|
|
|
| def extract_answer(response: str) -> str: |
| """ |
| 从模型响应中提取答案索引 |
| |
| Args: |
| response: 模型的原始响应 |
| |
| Returns: |
| "0", "1", "2" 或 ""（提取失败） |
| """ |
| |
| response = response.strip() |
| if response in ["0", "1", "2"]: |
| return response |
|
|
| |
| match = re.search(r"[012]", response) |
| if match: |
| return match.group(0) |
|
|
| |
| return "" |
|
|
|
|
| def compute_score(reward_inputs: list[dict[str, Any]]) -> list[dict[str, float]]: |
| """ |
| 计算一批样本的得分 |
| |
| Args: |
| reward_inputs: 包含多个样本的列表，每个样本包含: |
| - response: 模型的响应 |
| - response_length: 响应长度 |
| - ground_truth: 正确答案 |
| |
| Returns: |
| 每个样本的得分字典列表，包含: |
| - overall: 总分（1.0表示正确，0.0表示错误） |
| - accuracy: 准确率（同overall，用于监控） |
| """ |
| scores = [] |
|
|
| for reward_input in reward_inputs: |
| response = reward_input.get("response", "") |
| ground_truth = reward_input.get("ground_truth", "") |
|
|
| |
| predicted = extract_answer(response) |
|
|
| |
| if predicted == ground_truth: |
| score = 1.0 |
| else: |
| score = 0.0 |
|
|
| |
| scores.append({"overall": score, "accuracy": score}) |
|
|
| return scores |
|
|
|
|
| |
| if __name__ == "__main__": |
| test_cases = [ |
| |
| {"response": "0", "response_length": 1, "ground_truth": "0"}, |
| {"response": "1", "response_length": 1, "ground_truth": "1"}, |
| {"response": "2", "response_length": 1, "ground_truth": "2"}, |
| |
| {"response": "The answer is 1", "response_length": 15, "ground_truth": "1"}, |
| {"response": "I choose option 2", "response_length": 18, "ground_truth": "2"}, |
| |
| {"response": "0", "response_length": 1, "ground_truth": "1"}, |
| {"response": "2", "response_length": 1, "ground_truth": "0"}, |
| |
| {"response": "I don't know", "response_length": 12, "ground_truth": "1"}, |
| {"response": "", "response_length": 0, "ground_truth": "2"}, |
| ] |
|
|
| scores = compute_score(test_cases) |
|
|
| print("Reward Function Test Results:") |
| print("=" * 60) |
| for i, (test, score) in enumerate(zip(test_cases, scores), 1): |
| print(f"{i}. Response: {test['response']!r}") |
| print(f" Ground Truth: {test['ground_truth']!r}") |
| print(f" Score: {score}") |
| print() |
|
|
