File size: 10,819 Bytes
4c30587 | 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 | """
Base classes for evaluation.
This module provides abstract interfaces for evaluation that can be
extended to support different datasets. The evaluator handles:
1. Building prompts from retrieved memories
2. Generating predictions using LLM
3. Computing evaluation metrics
"""
from abc import ABC, abstractmethod
from typing import List, Dict, Any, Tuple
from dataclasses import dataclass, field
import copy
import numpy as np
@dataclass
class EvalResult:
"""
Result of evaluating a single QA item.
This provides a unified result format across different datasets.
"""
qa_idx: int
question: str
ground_truth: str
prediction: str
retrieved_memories: List[str]
retrieved_indices: List[int]
f1_score: float = 0.0
llm_judge_score: float = 0.0
is_correct: bool = False
metadata: Dict[str, Any] = field(default_factory=dict)
@dataclass
class EvalSummary:
"""Summary statistics from evaluation."""
avg_f1: float = 0.0
avg_llm_judge: float = 0.0
num_samples: int = 0
num_correct: int = 0
accuracy: float = 0.0
results: List[EvalResult] = field(default_factory=list)
category_scores: Dict[Any, Dict[str, float]] = field(default_factory=dict)
class Evaluator(ABC):
"""
Abstract base class for evaluation.
Subclasses should implement methods for:
1. Filtering QA items (which ones to skip)
2. Building prompts from retrieved memories
3. Any dataset-specific evaluation logic
"""
def __init__(self, args):
"""
Initialize the evaluator.
Args:
args: Arguments containing model config, etc.
"""
self.args = args
@abstractmethod
def filter_qa_list(self, qa_list: List[Dict]) -> List[Tuple[int, Dict]]:
"""
Filter QA list to get valid items for evaluation.
Args:
qa_list: List of QA dicts
Returns:
List of (index, qa_dict) tuples for valid QA items
"""
pass
@abstractmethod
def build_prompt(self, question: str, retrieved_memories: List[str],
qa_item: Dict) -> str:
"""
Build evaluation prompt from question and retrieved memories.
Args:
question: The question to answer
retrieved_memories: List of retrieved memory texts
qa_item: Original QA dict (may contain extra info like date)
Returns:
Formatted prompt string
"""
pass
def sample_train_qa_list(self,
qa_list: List[Dict],
valid_qa: List[Tuple[int, Dict]],
conversation_id: str = None,
outer_epoch: int = 0,
inner_epoch: int = 0) -> List[Tuple[int, Dict]]:
"""
Optional training-only QA sampling hook.
This hook is used by trainer-side reward evaluation only and should not
affect eval-only / formal evaluation paths unless explicitly invoked.
"""
return valid_qa
def get_ground_truth(self, qa_item: Dict) -> str:
"""
Extract ground truth answer from QA item.
Override in subclasses if needed.
Args:
qa_item: QA dict
Returns:
Ground truth answer string
"""
return str(qa_item.get('answer', ''))
def prepare_eval_args(self) -> Any:
"""
Prepare arguments for evaluation LLM calls.
Override in subclasses to customize.
Returns:
Modified args for evaluation
"""
eval_args = copy.deepcopy(self.args)
eval_args.max_new_tokens = 32
eval_args.temperature = 0.0
eval_args.batch_size = 32
return eval_args
def compute_f1(self, prediction: str, ground_truth: str, qa_item: Dict = None) -> float:
"""
Compute F1 score between prediction and ground truth.
Subclasses may override to use category-specific F1 methods.
Args:
prediction: Model prediction
ground_truth: Ground truth answer
qa_item: Original QA dict (for category-specific handling)
Returns:
F1 score (0-1)
"""
from eval_utils import f1_score
return f1_score(prediction, ground_truth)
def build_judge_prompt(self, question: str, ground_truth, prediction: str, qa_item: Dict) -> str:
"""
Build LLM judge prompt for evaluation.
Subclasses can override to provide task-specific judging criteria.
"""
from prompts.prompt_pool import LLM_JUDGE_GENERAL_PROMPT
if isinstance(ground_truth, list):
ground_truth_str = ", ".join(str(ans) for ans in ground_truth)
else:
ground_truth_str = str(ground_truth)
return LLM_JUDGE_GENERAL_PROMPT.format(
question=question,
ground_truth=ground_truth_str,
model_answer=prediction
)
def evaluate(self, qa_list: List[Dict], memory_bank: Any,
conversation_id: str = None,
collect_cases: bool = False) -> EvalSummary:
"""
Evaluate QA performance on a list of QA items.
This is the main evaluation entry point. It handles:
1. Filtering valid QA items
2. Retrieving memories for each question
3. Building prompts and calling LLM
4. Computing metrics
Args:
qa_list: List of QA dicts
memory_bank: MemoryBank to retrieve from
conversation_id: Optional ID for tracking
collect_cases: Whether to collect detailed cases
Returns:
EvalSummary with results and statistics
"""
from llm_utils import get_llm_response
from rag_utils import get_embeddings
# Filter valid QA items
valid_qa = self.filter_qa_list(qa_list)
if not valid_qa:
return EvalSummary()
eval_args = self.prepare_eval_args()
# Collect questions for batch embedding
questions = [qa['question'] for _, qa in valid_qa]
# Batch compute question embeddings
q_embeddings = get_embeddings(
self.args.retriever,
questions,
'query'
)
# Build task args for parallel LLM calls
task_args = []
retrieval_info = {} # qa_idx -> (retrieved_mems, retrieved_indices)
for idx, (qa_idx, qa) in enumerate(valid_qa):
question = qa['question']
q_embedding = q_embeddings[idx]
# Retrieve from memory bank
top_k = getattr(self.args, "mem_top_k_eval", None)
if top_k is None:
top_k = getattr(self.args, "mem_top_k", 5)
retrieved_mems, retrieved_indices = memory_bank.retrieve(
q_embedding, top_k=top_k, use_state_encoder=False
)
retrieval_info[qa_idx] = (retrieved_mems, list(retrieved_indices))
# Build prompt
prompt = self.build_prompt(question, retrieved_mems, qa)
task_args.append((qa_idx, prompt, eval_args))
if not task_args:
return EvalSummary()
# Call LLM in parallel
ret = get_llm_response(args=eval_args, task_args=task_args)
# Compute scores
results = []
f1_scores = []
for i, response, _, success in ret:
qa = qa_list[i]
ground_truth = self.get_ground_truth(qa)
prediction = response.strip() if success else ""
f1 = self.compute_f1(prediction, ground_truth)
f1_scores.append(f1)
retrieved_mems, retrieved_indices = retrieval_info.get(i, ([], []))
result = EvalResult(
qa_idx=i,
question=qa['question'],
ground_truth=ground_truth,
prediction=prediction,
retrieved_memories=retrieved_mems,
retrieved_indices=retrieved_indices,
f1_score=f1,
metadata=self._get_result_metadata(qa)
)
results.append(result)
# Compute summary statistics
avg_f1 = np.mean(f1_scores) if f1_scores else 0.0
return EvalSummary(
avg_f1=avg_f1,
num_samples=len(results),
results=results
)
def _get_result_metadata(self, qa: Dict) -> Dict[str, Any]:
"""
Extract metadata from QA item for result.
Override in subclasses to add dataset-specific metadata.
Args:
qa: QA dict
Returns:
Metadata dict
"""
return {}
def run_llm_judge(self, results: List[EvalResult]) -> List[float]:
"""
Run LLM judge on evaluation results.
Args:
results: List of EvalResult objects
Returns:
List of LLM judge scores (0.0, 0.5, or 1.0)
"""
from eval_utils import llm_judge
from prompts.prompt_pool import LLM_JUDGE_GENERAL_PROMPT
if not results:
return []
eval_args = self.prepare_eval_args()
judge_task_args = []
for result in results:
prompt = LLM_JUDGE_GENERAL_PROMPT.format(
question=result.question,
ground_truth=result.ground_truth,
model_answer=result.prediction
)
judge_task_args.append((result.qa_idx, prompt, eval_args))
scores = llm_judge(task_args=judge_task_args, args=eval_args)
# Update results with LLM judge scores
for result, score in zip(results, scores):
result.llm_judge_score = score
result.is_correct = (score == 1.0)
return scores
# Global registry for evaluators
_EVALUATOR_REGISTRY: Dict[str, type] = {}
def register_evaluator(name: str):
"""
Decorator to register an evaluator class.
Usage:
@register_evaluator("my_dataset")
class MyEvaluator(Evaluator):
...
"""
def decorator(cls):
_EVALUATOR_REGISTRY[name] = cls
return cls
return decorator
def get_evaluator(name: str, args) -> Evaluator:
"""
Get an evaluator instance by name.
Args:
name: Registered evaluator name
args: Arguments to pass to evaluator constructor
Returns:
Evaluator instance
"""
if name not in _EVALUATOR_REGISTRY:
raise ValueError(f"Unknown evaluator: {name}. Available: {list(_EVALUATOR_REGISTRY.keys())}")
return _EVALUATOR_REGISTRY[name](args)
def list_evaluators() -> List[str]:
"""List all registered evaluator names."""
return list(_EVALUATOR_REGISTRY.keys())
|