You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
hello-agents/code/chapter12/09_data_generation_win_rate.py

171 lines
5.0 KiB
Python

6 months ago
"""
第十二章示例9Win Rate评估
对应文档12.4.4 Win Rate评估
这个示例展示如何使用Win Rate评估生成的AIME题目质量
Win Rate评估通过对比生成题目和真题评估生成质量
- Win Rate = 50%生成质量与真题相当理想情况
- Win Rate > 50%生成质量优于真题可能是评估偏差
- Win Rate < 50%生成质量低于真题需要改进
"""
import sys
import os
import json
# 添加HelloAgents路径
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "HelloAgents"))
from hello_agents import HelloAgentsLLM
from hello_agents.evaluation import WinRateEvaluator, AIDataset
# 1. 准备生成的题目数据
generated_problems = [
{
"problem_id": "generated_001",
"problem": "Find the number of positive integers $n$ such that $n^2 + 19n + 92$ is a perfect square.",
"answer": "4"
},
{
"problem_id": "generated_002",
"problem": "In triangle $ABC$, $AB = 13$, $BC = 14$, and $CA = 15$. Find the area of the triangle.",
"answer": "84"
},
{
"problem_id": "generated_003",
"problem": "How many positive integers less than 1000 are divisible by 7 but not by 11?",
"answer": "129"
}
]
# 2. 加载参考数据集AIME真题
print("="*60)
print("Win Rate评估")
print("="*60)
print("\n加载参考数据集...")
dataset = AIDataset()
reference_problems = dataset.load()
print(f"✅ 已加载 {len(reference_problems)} 道AIME真题")
# 3. 创建Win Rate评估器
llm = HelloAgentsLLM(model_name="gpt-4o")
evaluator = WinRateEvaluator(
llm=llm,
reference_problems=reference_problems
)
# 4. 运行Win Rate评估
print(f"\n开始Win Rate评估...")
print(f" 生成题目数: {len(generated_problems)}")
print(f" 对比数量: 20")
results = evaluator.evaluate(
generated_problems=generated_problems,
num_comparisons=20 # 进行20次对比
)
# 5. 显示评估结果
print("\n" + "="*60)
print("评估结果")
print("="*60)
print(f"\nWin Rate: {results['win_rate']:.2%}")
print(f"Tie Rate: {results['tie_rate']:.2%}")
print(f"Loss Rate: {results['loss_rate']:.2%}")
print(f"\n详细统计:")
print(f" 总对比数: {results['total_comparisons']}")
print(f" 生成题目胜: {results['wins']}")
print(f" 平局: {results['ties']}")
print(f" 真题胜: {results['losses']}")
# 6. 质量评估
print(f"\n质量评估:")
win_rate = results['win_rate']
if 0.45 <= win_rate <= 0.55:
print("✅ 优秀 - 生成质量接近AIME真题水平")
elif 0.35 <= win_rate < 0.45:
print("⚠️ 良好 - 生成质量可用,但略低于真题")
elif 0.25 <= win_rate < 0.35:
print("⚠️ 一般 - 生成质量一般,需要改进")
else:
print("❌ 较差 - 生成质量差,需要大幅改进")
# 7. 查看部分对比详情
print("\n" + "="*60)
print("对比详情前5个")
print("="*60)
for i, comparison in enumerate(results['comparisons'][:5], 1):
print(f"\n对比 {i}:")
print(f" 生成题目: {comparison['generated_problem'][:60]}...")
print(f" 真题: {comparison['reference_problem'][:60]}...")
print(f" 结果: {comparison['result']}")
if 'reason' in comparison:
print(f" 理由: {comparison['reason'][:100]}...")
# 8. 保存评估结果
output_file = "./evaluation_results/win_rate_results.json"
os.makedirs(os.path.dirname(output_file), exist_ok=True)
with open(output_file, 'w', encoding='utf-8') as f:
json.dump(results, f, indent=2, ensure_ascii=False)
print(f"\n✅ 评估结果已保存到 {output_file}")
# 运行输出示例:
# ============================================================
# Win Rate评估
# ============================================================
#
# 加载参考数据集...
# ✅ 已加载 963 道AIME真题
#
# 开始Win Rate评估...
# 生成题目数: 3
# 对比数量: 20
#
# Win Rate评估: 100%|██████████| 20/20 [01:00<00:00, 3.01s/对比]
#
# ============================================================
# 评估结果
# ============================================================
#
# Win Rate: 45.00%
# Tie Rate: 10.00%
# Loss Rate: 45.00%
#
# 详细统计:
# 总对比数: 20
# 生成题目胜: 9
# 平局: 2
# 真题胜: 9
#
# 质量评估:
# ✅ 优秀 - 生成质量接近AIME真题水平
#
# ============================================================
# 对比详情前5个
# ============================================================
#
# 对比 1:
# 生成题目: Find the number of positive integers $n$ such that $n^2 + 19...
# 真题: Let $N$ be the number of consecutive $0$'s at the right end...
# 结果: generated
# 理由: The generated problem has a clearer problem statement and a mo...
#
# 对比 2:
# 生成题目: In triangle $ABC$, $AB = 13$, $BC = 14$, and $CA = 15$. F...
# 真题: Find the number of ordered pairs $(m,n)$ of positive integers...
# 结果: reference
# 理由: The reference problem is more challenging and requires deeper...
#
# ...
#
# ✅ 评估结果已保存到 ./evaluation_results/win_rate_results.json