Spaces:
Running
Running
File size: 19,153 Bytes
1ea875f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 |
# 文件路径: evaluation/evaluation_framework.py
"""
GitHub Agent 完整评估框架
四层评估架构 + 数据路由引擎
Author: Dexter
Date: 2025-01-27
注意: 数据模型已拆分到 models.py,数据路由已拆分到 data_router.py
此文件保留核心评估引擎逻辑,并重新导出所有符号保持向后兼容
"""
import json
import os
import re
from typing import List, Dict, Any
from datetime import datetime
# 重新导出所有模型(保持向后兼容)
from evaluation.models import (
EvaluationLayer,
DataQualityTier,
QueryRewriteMetrics,
RetrievalMetrics,
GenerationMetrics,
AgenticMetrics,
EvaluationResult,
)
from evaluation.data_router import DataRoutingEngine
# ============================================================================
# 评估引擎核心逻辑
# ============================================================================
class EvaluationEngine:
"""评估引擎 - 负责多层面打分"""
def __init__(
self,
llm_client=None,
golden_dataset_path: str = "evaluation/golden_dataset.json",
model_name: str = None
):
self.llm_client = llm_client
self.model_name = model_name or "gpt-4o-mini" # 默认使用轻量模型
self.golden_dataset = self._load_golden_dataset(golden_dataset_path)
def _load_golden_dataset(self, path: str) -> List[Dict]:
"""加载黄金数据集"""
if not os.path.exists(path):
print(f"⚠️ Golden dataset not found at {path}")
return []
with open(path, 'r', encoding='utf-8') as f:
return json.load(f)
async def evaluate_query_rewrite(
self,
original_query: str,
rewritten_query: str,
language_detected: str
) -> QueryRewriteMetrics:
"""
评估查询重写质量
指标:
- keyword_coverage: 重写后的关键词是否覆盖了原Query的核心概念?
- semantic_preservation: 语义是否保留?
- diversity_score: 关键词多样性
"""
# 简化版: 使用关键词匹配
original_tokens = set(original_query.lower().split())
rewritten_tokens = set(rewritten_query.lower().split())
# 关键词覆盖度: 原Query的关键词有多少在重写中保留
if original_tokens:
coverage = len(original_tokens & rewritten_tokens) / len(original_tokens)
else:
coverage = 0.0
# 多样性: 重写后的关键词数量越多、越不重复,分数越高
unique_ratio = len(rewritten_tokens) / max(len(original_tokens), 1)
diversity = min(1.0, unique_ratio)
# 语义保留度 (简化版本: 假设如果覆盖度高就认为语义保留良好)
semantic_preservation = min(1.0, coverage + 0.2) # 基础分+覆盖度加分
return QueryRewriteMetrics(
original_query=original_query,
rewritten_query=rewritten_query,
language_detected=language_detected,
keyword_coverage=coverage,
semantic_preservation=semantic_preservation,
diversity_score=diversity
)
async def evaluate_retrieval(
self,
query: str,
retrieved_files: List[str],
ground_truth_files: List[str],
top_k: int = 5,
retrieval_latency_ms: float = 0,
vector_scores: List[float] = None,
bm25_scores: List[float] = None
) -> RetrievalMetrics:
"""
评估检索层质量
指标:
- hit_rate: 是否找到了任何正确的文件?
- recall_at_k: 前K个中有多少是正确的?
- precision_at_k: 返回的文件中有多少是正确的?
- mrr: 第一个正确结果的排名倒数
"""
retrieved_set = set(retrieved_files[:top_k])
ground_truth_set = set(ground_truth_files)
# Hit rate: 是否有交集
hit_rate = 1.0 if retrieved_set & ground_truth_set else 0.0
# Recall@K: 找到的正确结果数 / 正确结果总数
correct_count = len(retrieved_set & ground_truth_set)
recall = correct_count / len(ground_truth_set) if ground_truth_set else 0.0
# Precision@K: 找到的正确结果数 / 返回的结果总数
precision = correct_count / len(retrieved_set) if retrieved_set else 0.0
# MRR: 第一个正确结果的倒数排名
mrr = 0.0
for i, file in enumerate(retrieved_files[:top_k], 1):
if file in ground_truth_set:
mrr = 1.0 / i
break
# Context Relevance: 简化版 - 假设Precision反映了相关性
context_relevance = precision
# Chunk Integrity: 简化版 - 假设没有太多文件就认为完整度高
chunk_integrity = min(1.0, 1.0 / len(retrieved_set)) if retrieved_set else 0.0
vector_avg = sum(vector_scores) / len(vector_scores) if vector_scores else 0.0
bm25_avg = sum(bm25_scores) / len(bm25_scores) if bm25_scores else 0.0
return RetrievalMetrics(
query=query,
top_k=top_k,
hit_rate=hit_rate,
recall_at_k=recall,
precision_at_k=precision,
mrr=mrr,
context_relevance=context_relevance,
chunk_integrity=chunk_integrity,
retrieval_latency_ms=retrieval_latency_ms,
vector_score_avg=vector_avg,
bm25_score_avg=bm25_avg,
retrieved_files=retrieved_files,
ground_truth_files=ground_truth_files
)
async def evaluate_generation(
self,
query: str,
retrieved_context: str,
generated_answer: str,
ground_truth_answer: str = "",
generation_latency_ms: float = 0,
token_usage: Dict[str, int] = None
) -> GenerationMetrics:
"""
评估生成层质量
指标:
- faithfulness: 回答是否严格基于Context?
- answer_relevance: 回答是否回答了问题?
- answer_completeness: 回答是否足够完整?
- code_correctness: 生成的代码是否正确?
"""
# 1. Faithfulness: 使用LLM-as-Judge进行幻觉检测
faithfulness = await self._judge_faithfulness(
retrieved_context,
generated_answer
)
# 2. Answer Relevance: 回答和问题的相似度
answer_relevance = await self._judge_answer_relevance(
query,
generated_answer
)
# 3. Answer Completeness: 简化版 - 通过长度和结构判断
completeness = self._judge_completeness(
generated_answer,
ground_truth_answer
)
# 4. Code Correctness: 使用AST检查代码块
code_samples = self._extract_code_blocks(generated_answer)
code_correctness = self._check_code_correctness(code_samples)
metrics = GenerationMetrics(
query=query,
retrieved_context=retrieved_context,
generated_answer=generated_answer,
ground_truth_answer=ground_truth_answer,
faithfulness=faithfulness,
answer_relevance=answer_relevance,
answer_completeness=completeness,
code_correctness=code_correctness,
generated_code_samples=code_samples,
generation_latency_ms=generation_latency_ms,
token_usage=token_usage or {"input": 0, "output": 0}
)
return metrics
async def _judge_faithfulness(self, context: str, answer: str) -> float:
"""
LLM-as-Judge: 判断回答是否由Context支撑
返回 0-1 的分数
注意:Faithfulness 判断的是"回答中的信息是否能从 Context 中找到依据"
而不是"回答是否完全复制 Context 内容"
"""
if not self.llm_client:
# 简化版: 如果没有LLM客户端,使用启发式方法
# 统计Answer中的关键词有多少出现在Context中
context_lower = context.lower()
answer_words = set(answer.lower().split())
# 过滤掉常见停用词
stop_words = {'the', 'a', 'an', 'is', 'are', 'was', 'were', 'be', 'been',
'being', 'have', 'has', 'had', 'do', 'does', 'did', 'will',
'would', 'could', 'should', 'may', 'might', 'must', 'shall',
'can', 'need', 'dare', 'ought', 'used', 'to', 'of', 'in',
'for', 'on', 'with', 'at', 'by', 'from', 'as', 'into', 'that',
'which', 'who', 'whom', 'this', 'these', 'those', 'it', 'its'}
meaningful_words = answer_words - stop_words
if not meaningful_words:
return 0.7 # 没有有意义的词,给默认分
# 计算答案中有多少有意义的词出现在Context中
found_count = sum(1 for word in meaningful_words if word in context_lower)
overlap = found_count / len(meaningful_words)
return min(1.0, overlap + 0.2) # 给一定的基础分
# 智能截取 Context:提取与 Answer 相关的部分
# 如果 Context 太长,优先包含 Answer 中提到的关键词附近的内容
max_context_len = 6000 # 增加到 6000 字符
if len(context) > max_context_len:
# 尝试找到 Answer 中提到的关键文件/函数名
import re
# 提取 Answer 中可能的文件路径或函数名
patterns = re.findall(r'[a-zA-Z_][a-zA-Z0-9_]*(?:\.[a-zA-Z_][a-zA-Z0-9_]*)*', answer[:500])
important_terms = [p for p in patterns if len(p) > 3][:5] # 取前5个重要词
# 优先截取包含这些词的部分
context_parts = []
remaining = max_context_len
for term in important_terms:
idx = context.find(term)
if idx != -1 and remaining > 0:
start = max(0, idx - 300)
end = min(len(context), idx + 700)
snippet = context[start:end]
if snippet not in ''.join(context_parts):
context_parts.append(snippet)
remaining -= len(snippet)
# 如果没找到相关部分,还是用前 6000 字符
if context_parts:
truncated_context = "\n...\n".join(context_parts)
else:
truncated_context = context[:max_context_len]
else:
truncated_context = context
# 改进的 Prompt:更明确定义 Faithfulness
prompt = f"""Evaluate the FAITHFULNESS of the answer to the given context.
FAITHFULNESS means: The claims and information in the answer can be verified from or are consistent with the context.
- Score HIGH (0.7-1.0) if the answer correctly identifies or explains concepts that ARE in the context
- Score MEDIUM (0.4-0.7) if the answer is partially supported but makes some unsupported claims
- Score LOW (0.0-0.4) if the answer contradicts the context or makes completely unsupported claims
NOTE: If the answer says "X is not in the context" and X is indeed not shown, that's a FAITHFUL statement (score 0.7+)
NOTE: If the answer correctly identifies WHERE something is defined based on imports/references in context, that's FAITHFUL
[Context]
{truncated_context}
[Answer]
{answer[:1500]}
SCORE (0.0-1.0):"""
try:
response = await self.llm_client.chat.completions.create(
model=self.model_name,
messages=[{"role": "user", "content": prompt}],
temperature=0.1,
max_tokens=10
)
score_str = response.choices[0].message.content.strip()
# 提取数字(处理可能的额外文本)
import re
match = re.search(r'(\d+\.?\d*)', score_str)
if match:
score = float(match.group(1))
else:
score = float(score_str)
return min(1.0, max(0.0, score))
except Exception as e:
print(f"⚠️ Faithfulness judgment failed: {e}")
return 0.5
async def _judge_answer_relevance(self, query: str, answer: str) -> float:
"""判断回答与问题的相关性"""
if not self.llm_client:
# 简化版: 使用关键词重叠度
query_words = set(query.lower().split())
answer_words = set(answer.lower().split())
overlap = len(query_words & answer_words) / max(len(query_words), 1)
return min(1.0, overlap + 0.3) # 基础分0.3+重叠度
prompt = f"""
Does the answer address the query?
[Query]
{query}
[Answer]
{answer[:1000]}
Score (0.0-1.0):
"""
try:
response = await self.llm_client.chat.completions.create(
model=self.model_name,
messages=[{"role": "user", "content": prompt}],
temperature=0.1,
max_tokens=10
)
score = float(response.choices[0].message.content.strip())
return min(1.0, max(0.0, score))
except:
return 0.5
def _judge_completeness(self, generated_answer: str, ground_truth: str = "") -> float:
"""判断回答的完整性"""
# 简化版: 根据长度和结构
if len(generated_answer) < 50:
return 0.3
elif len(generated_answer) < 200:
return 0.6
else:
return 0.9
def _extract_code_blocks(self, text: str) -> List[str]:
"""从文本中提取代码块"""
import re
code_pattern = r'```[\w]*\n(.*?)\n```'
matches = re.findall(code_pattern, text, re.DOTALL)
return matches
def _check_code_correctness(self, code_samples: List[str]) -> float:
"""检查代码是否有语法错误"""
if not code_samples:
return 1.0 # 没有代码就认为正确
import ast
correct_count = 0
for code in code_samples:
try:
ast.parse(code)
correct_count += 1
except SyntaxError:
pass
return correct_count / len(code_samples)
async def evaluate_agentic(
self,
query: str,
tool_calls: List[Dict[str, Any]],
success: bool,
steps_taken: int = 0,
end_to_end_latency_ms: float = 0
) -> AgenticMetrics:
"""
评估Agent的决策和行为
"""
# Tool Selection Accuracy: 工具选择是否正确?
tool_selection_accuracy = 1.0 if success else 0.5
# Tool Parameter Correctness: 参数是否正确传递?
tool_param_correctness = 1.0 if all(
tc.get("success", False) for tc in tool_calls
) else 0.5
# 计算冗余步骤
unnecessary_steps = 0
backtrack_count = 0
# 简化版: 如果有重复的工具调用则视为冗余
tool_call_signatures = [tc.get("name", "") for tc in tool_calls]
for i, sig in enumerate(tool_call_signatures):
if i > 0 and sig == tool_call_signatures[i-1]:
unnecessary_steps += 1
return AgenticMetrics(
query=query,
tool_calls=tool_calls,
tool_selection_accuracy=tool_selection_accuracy,
tool_parameter_correctness=tool_param_correctness,
steps_taken=steps_taken,
unnecessary_steps=unnecessary_steps,
backtrack_count=backtrack_count,
success=success,
end_to_end_latency_ms=end_to_end_latency_ms
)
def get_statistics(self) -> Dict[str, Any]:
"""
获取评估统计信息
Returns:
包含 total_evaluations, average_score, quality_distribution, top_issues 的字典
"""
# 从 eval_results.jsonl 读取评估结果
eval_results_path = "evaluation/sft_data/eval_results.jsonl"
stats = {
"total_evaluations": 0,
"average_score": 0.0,
"quality_distribution": {
"gold": 0,
"silver": 0,
"bronze": 0,
"rejected": 0
},
"top_issues": []
}
if not os.path.exists(eval_results_path):
return stats
# 读取和分析评估结果
scores = []
issues = {}
try:
with open(eval_results_path, 'r', encoding='utf-8') as f:
for line in f:
try:
result = json.loads(line)
stats["total_evaluations"] += 1
# 收集得分
score = result.get("overall_score", 0)
scores.append(score)
# 统计质量分布
tier = result.get("data_quality_tier", "bronze")
if tier in stats["quality_distribution"]:
stats["quality_distribution"][tier] += 1
# 收集常见问题 (假设记录在 notes 或 error_message 中)
note = result.get("notes", "") or result.get("error_message", "")
if note:
issues[note] = issues.get(note, 0) + 1
except json.JSONDecodeError:
continue
except Exception as e:
print(f"⚠️ Error reading eval results: {e}")
# 计算平均分
if scores:
stats["average_score"] = sum(scores) / len(scores)
# 获取前5个常见问题
if issues:
stats["top_issues"] = [
{"issue": issue, "count": count}
for issue, count in sorted(issues.items(), key=lambda x: x[1], reverse=True)[:5]
]
return stats
# ============================================================================
# __all__ 导出列表(保持向后兼容)
# ============================================================================
__all__ = [
# 枚举
"EvaluationLayer",
"DataQualityTier",
# 数据模型
"QueryRewriteMetrics",
"RetrievalMetrics",
"GenerationMetrics",
"AgenticMetrics",
"EvaluationResult",
# 引擎
"EvaluationEngine",
"DataRoutingEngine",
]
|