eval_hellaswag.py

import json
import re
import argparse


# 字母到数字的映射
letter_to_number = {
    "A": "0",
    "B": "1",
    "C": "2",
    "D": "3"
}

# 函数：从text中提取模型预测的答案
def extract_predicted_answer(text):
    # 匹配类似 "A", "B", "C", "D" 的答案
    match = re.search(r'\b[A-D]\b', text)
    if match:
        return match.group(0)
    return None

# 函数：计算正确率
def calculate_accuracy(file_path):
    total = 0
    correct = 0

    # 逐行读取JSONL文件
    with open(file_path, 'r', encoding='utf-8') as file:
        for line in file:
            item = json.loads(line)
            total += 1
            pred_letter = extract_predicted_answer(item['text'])  # 从text中提取答案
            true_answer = item['answer']                  # 正确答案
            if pred_letter:
                pred = letter_to_number[pred_letter]  # 映射到数字 (1-4)
                true_answer = str(item['answer'])     # 转为字符串，避免类型问题
                if pred == true_answer:
                    correct += 1

    # 计算正确率
    accuracy = (correct / total) * 100

    # 打印结果
    print(f"总样本数: {total}")
    print(f"正确样本数: {correct}")
    print(f"错误样本数: {total - correct}")
    print(f"正确率: {accuracy:.2f}%")

# 主程序
if __name__ == "__main__":
    # 使用 argparse 获取命令行参数
    parser = argparse.ArgumentParser(description="Evaluate model accuracy from JSONL file.")
    parser.add_argument("file_path", type=str, help="Path to the JSONL file containing inference results.")
    args = parser.parse_args()

    # 调用计算函数
    calculate_accuracy(args.file_path)