File size: 16,220 Bytes
4e909c7 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 |
"""
基础模型 vs 微调模型对比测试 - 增强版
重点测试项目特定知识
核心改进:
1. 增强 Repo-Specific 评估权重
2. 提高对 project name, file/class name 提及的奖励
"""
import os
import json
import yaml
import torch
from pathlib import Path
from transformers import AutoTokenizer, AutoModelForCausalLM
from typing import List, Dict
from dataclasses import dataclass, asdict
from datetime import datetime
@dataclass
class TestCase:
"""测试用例"""
type: str # repo_specific, code_specific, general
question: str
category: str # overview, architecture, implementation, etc.
reference_files: List[str] = None # 参考的代码文件
class ModelComparator:
"""模型对比器"""
def __init__(self, config_path: str = "../config/default_config.yaml"):
with open(config_path, 'r', encoding='utf-8') as f:
self.config = yaml.safe_load(f)
# 加载分析数据
analysis_path = "./data/repository_analysis.json"
try:
with open(analysis_path, 'r', encoding='utf-8') as f:
self.analysis_data = json.load(f)
except FileNotFoundError:
print(f"❌ 警告: 找不到分析文件 {analysis_path}。使用默认空数据。")
self.analysis_data = {'code_elements': [], 'project_context': {'project_name': 'Laddr'}}
self.project_name = self.analysis_data.get('project_context', {}).get('project_name', 'Laddr')
self.code_elements = self.analysis_data.get('code_elements', [])
self.base_model = None
self.base_tokenizer = None
self.finetuned_model = None
self.finetuned_tokenizer = None
def load_models(self):
"""加载基础模型和微调模型"""
base_model_path = self.config['model']['base_model']
finetuned_model_path = Path(self.config['training']['output_dir']) / "merged_model"
# 检查 merged model 是否存在
if not finetuned_model_path.exists():
print(f"❌ 错误: 找不到合并后的模型 {finetuned_model_path}。请先运行 merge_lora.py。")
return
print("Loading base model...")
self.base_tokenizer = AutoTokenizer.from_pretrained(
base_model_path,
trust_remote_code=True
)
self.base_model = AutoModelForCausalLM.from_pretrained(
base_model_path,
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True
)
self.base_model.eval()
print("Loading fine-tuned model...")
self.finetuned_tokenizer = AutoTokenizer.from_pretrained(
str(finetuned_model_path),
trust_remote_code=True
)
self.finetuned_model = AutoModelForCausalLM.from_pretrained(
str(finetuned_model_path),
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True
)
self.finetuned_model.eval()
print("Models loaded successfully")
def generate_response(self, model, tokenizer, question: str, max_new_tokens: int = 1024) -> str:
"""生成回答"""
messages = [
# Qwen 默认无 system prompt
{"role": "user", "content": question}
]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=max_new_tokens,
temperature=0.7,
do_sample=True,
top_p=0.9,
pad_token_id=tokenizer.pad_token_id
)
# Qwen 模型会生成 <|im_start|>assistant\n 和答案
response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:],
skip_special_tokens=True)
# 移除可能的角色标识
if response.startswith("assistant"):
response = response[len("assistant"):].strip()
return response.strip()
def create_test_cases(self) -> List[TestCase]:
"""创建测试用例 - 基于实际代码内容"""
test_cases = []
# 1. 项目概览问题 (使用项目实际信息)
test_cases.extend([
TestCase(
type="repo_specific",
question=f"{self.project_name} 项目的主要功能是什么?",
category="overview"
),
TestCase(
type="repo_specific",
question=f"请介绍 {self.project_name} 的架构设计。",
category="architecture"
),
TestCase(
type="repo_specific",
question=f"{self.project_name} 中有哪些核心模块?",
category="modules"
),
])
# 2. 具体代码相关问题
# 找几个关键函数/类
key_elements = sorted(self.code_elements,
key=lambda x: x.get('complexity', 0),
reverse=True)[:5]
for element in key_elements:
name = element['name']
filepath = element['filepath']
test_cases.append(TestCase(
type="code_specific",
question=f"{self.project_name} 中 `{name}` 的作用是什么?",
category="code_understanding",
reference_files=[filepath]
))
if element['type'] in ['function', 'method'] and not name.startswith('_'):
test_cases.append(TestCase(
type="code_specific",
question=f"如何使用 {self.project_name} 的 `{name}`?",
category="code_usage",
reference_files=[filepath]
))
# 增加代码定位问题
test_cases.append(TestCase(
type="code_location",
question=f"`{name}` 在 {self.project_name} 哪个文件里?",
category="code_location",
reference_files=[filepath]
))
# 3. 通用软件工程问题 (对照组)
test_cases.extend([
TestCase(
type="general",
question="什么是代码重构?",
category="general_knowledge"
),
TestCase(
type="general",
question="如何设计一个可扩展的软件架构?",
category="general_knowledge"
),
])
return test_cases
def evaluate_response(self, question: str, response: str, test_type: str) -> Dict:
"""评估回答质量 - 增强版"""
score = 0.0
# 1. 长度合理性 (50-1000字符)
length = len(response)
# 避免回答太短(幻觉或拒绝回答)或太长(填充无关信息)
if 50 <= length <= 1000:
score += 0.3
elif 20 <= length < 50 or length > 1000:
score += 0.1
# 2. 提及项目名称 (repo_specific, code_specific, code_location)
has_repo_mention = self.project_name.lower() in response.lower()
if test_type in ["repo_specific", "code_specific", "code_location"]:
if has_repo_mention:
score += 0.25 # 基础分数
# 检查是否提及问题中的核心元素 (如函数名、类名)
question_words = [q.strip('`?!.,') for q in question.split() if q.strip('`?!.,') not in ['Laddr', self.project_name, '请解释', '在', '如何', '中', '的', '作用', '是什么', '哪个', '文件', '里', '使用', '函数', '方法', '类']]
# 检查是否提到任何代码元素
code_element_mentions = 0
for elem in self.code_elements[:50]: # 只检查前50个,防止计算量过大
if f"`{elem['name']}`".lower() in response.lower():
code_element_mentions += 1
if code_element_mentions > 0:
score += 0.25 # 提及代码元素,强烈暗示学习到了新知识
# 检查是否提及文件路径 (code_location or code_specific)
if test_type in ["code_location", "code_specific"] and any(f for f in self.code_elements[:50] if f['filepath'] in response):
score += 0.1 # 提到文件路径,也是强相关性
# 3. 具体性 (vs 模糊回答)
uncertainty_phrases = ['可能', '大概', '或许', '不确定', '无法确认', '没有找到', '没有广泛知名']
has_uncertainty = any(phrase in response for phrase in uncertainty_phrases)
# 如果是项目特定/代码特定的问题,避免不确定性
if test_type in ["repo_specific", "code_specific", "code_location"] and not has_uncertainty:
score += 0.1
# 如果是通用问题,不惩罚不确定性
# 4. 代码块/路径 (code_usage, code_location)
if test_type in ["code_specific", "code_location"]:
if '```' in response or '`' in response: # 代码块或反引号标记
score += 0.1
# 标准化到0-1
score = min(1.0, score)
# 针对通用问题 (对照组) 设定上限
if test_type == "general":
score = min(score, 0.4) # 通用问题得分不应超过 0.4,因为我们主要优化的是特定知识
return {
'score': score,
'length': length,
'has_repo_mention': has_repo_mention,
'has_uncertainty': has_uncertainty,
'specifics_count': response.count('`'),
'quality': 'excellent' if score >= 0.8 else 'good' if score >= 0.6 else 'fair' if score >= 0.3 else 'poor'
}
def run_comparison(self):
"""运行对比测试"""
if self.finetuned_model is None:
print("无法运行对比:模型加载失败。")
return None
print(f"\nRunning comparison tests for {self.project_name}...")
test_cases = self.create_test_cases()
results = []
for i, test in enumerate(test_cases, 1):
print(f"\n[{i}/{len(test_cases)}] Testing: {test.question}")
# 基础模型回答
print(" Generating base model response...")
base_response = self.generate_response(
self.base_model,
self.base_tokenizer,
test.question
)
# 微调模型回答
print(" Generating fine-tuned model response...")
finetuned_response = self.generate_response(
self.finetuned_model,
self.finetuned_tokenizer,
test.question
)
# 评估
base_eval = self.evaluate_response(test.question, base_response, test.type)
finetuned_eval = self.evaluate_response(test.question, finetuned_response, test.type)
improvement = finetuned_eval['score'] - base_eval['score']
results.append({
'test_case': asdict(test),
'base_response': base_response,
'base_evaluation': base_eval,
'finetuned_response': finetuned_response,
'finetuned_evaluation': finetuned_eval,
'improvement': improvement
})
print(f" Base score: {base_eval['score']:.2f} | Fine-tuned score: {finetuned_eval['score']:.2f} | Improvement: {improvement:+.2f}")
return results
# ... (generate_report 函数不变)
def generate_report(self, results: List[Dict]):
"""生成对比报告"""
total_tests = len(results)
base_scores = [r['base_evaluation']['score'] for r in results]
finetuned_scores = [r['finetuned_evaluation']['score'] for r in results]
improvements = [r['improvement'] for r in results]
overall_base = sum(base_scores) / len(base_scores)
overall_finetuned = sum(finetuned_scores) / len(finetuned_scores)
overall_improvement = sum(improvements) / len(improvements)
# 按类型统计
type_stats = {}
for result in results:
test_type = result['test_case']['type']
if test_type not in type_stats:
type_stats[test_type] = {
'count': 0,
'base_scores': [],
'finetuned_scores': []
}
type_stats[test_type]['count'] += 1
type_stats[test_type]['base_scores'].append(result['base_evaluation']['score'])
type_stats[test_type]['finetuned_scores'].append(result['finetuned_evaluation']['score'])
type_statistics = {}
for test_type, stats in type_stats.items():
type_statistics[test_type] = {
'count': stats['count'],
'avg_base': sum(stats['base_scores']) / len(stats['base_scores']),
'avg_finetuned': sum(stats['finetuned_scores']) / len(stats['finetuned_scores']),
'avg_improvement': (sum(stats['finetuned_scores']) - sum(stats['base_scores'])) / len(stats['base_scores'])
}
# 判定
if overall_improvement > 0.15:
conclusion = "✓ 微调效果显著"
elif overall_improvement > 0.05:
conclusion = "⚠️ 微调有提升"
else:
conclusion = "✗ 微调效果不明显或失败"
report = {
'metadata': {
'repo_name': self.project_name,
'test_date': datetime.now().isoformat(),
'base_model': self.config['model']['base_model'],
'finetuned_model': str(Path(self.config['training']['output_dir']) / "merged_model")
},
'summary': {
'total_tests': total_tests,
'overall_base_score': overall_base,
'overall_finetuned_score': overall_finetuned,
'overall_improvement': overall_improvement,
'conclusion': conclusion
},
'type_statistics': type_statistics,
'detailed_results': results
}
# 保存报告
output_path = f"./comparison_report_{self.project_name}.json"
with open(output_path, 'w', encoding='utf-8') as f:
json.dump(report, f, indent=2, ensure_ascii=False)
print(f"\n{'='*60}")
print("Comparison Report")
print(f"{'='*60}")
print(f"Project: {self.project_name}")
print(f"Total Tests: {total_tests}")
print(f"Base Model Score: {overall_base:.3f}")
print(f"Fine-tuned Model Score: {overall_finetuned:.3f}")
print(f"Improvement: {overall_improvement:+.3f} ({overall_improvement/overall_base*100:+.1f}%)")
print(f"\nConclusion: {conclusion}")
print(f"\nDetailed report saved to: {output_path}")
print(f"{'='*60}")
return report
def main():
"""主函数"""
if 'CUDA_VISIBLE_DEVICES' not in os.environ:
os.environ['CUDA_VISIBLE_DEVICES'] = '0' # 确保单个 GPU 也能运行
print("="*60)
print("Base vs Fine-tuned Model Comparison (Enhanced Evaluation)")
print("="*60)
comparator = ModelComparator()
comparator.load_models()
results = comparator.run_comparison()
if results is not None:
report = comparator.generate_report(results)
print("\nComparison completed!")
if __name__ == "__main__":
main()
|