File size: 20,144 Bytes
66153d5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 | """
Phase 3 β Evaluation Framework
================================
Multi-metric evaluation pipeline:
β’ ROUGE (rouge1, rouge2, rougeL)
β’ BERTScore
β’ Faithfulness / hallucination detection
β’ LLM-as-Judge (model comparison)
β’ Knowledge gap detector (feeds Phase 4)
Usage:
python -m phase3_evaluation.evaluate --model ./merged --dataset squad --output ./evals
"""
from __future__ import annotations
import json
import re
from dataclasses import dataclass, field, asdict
from pathlib import Path
from typing import Optional
import torch
import typer
from rich.console import Console
from rich.table import Table
from rich.progress import track
from configs.settings import (
EVAL_DIR, ROUGE_TYPES, JUDGE_MODEL, CFG, HF_TOKEN
)
from utils.logger import logger
app = typer.Typer(help="Phase 3: Evaluation framework")
console = Console()
# βββββββββββββββββββββββββββββββββββββββββββββ
# Data models
# βββββββββββββββββββββββββββββββββββββββββββββ
@dataclass
class Sample:
id: str
context: str
question: str
reference: str # ground-truth answer
prediction: str = "" # filled during eval
@dataclass
class SampleScore:
sample_id: str
rouge1: float = 0.0
rouge2: float = 0.0
rougeL: float = 0.0
bertscore_f1: float = 0.0
faithfulness: float = 0.0 # 0β1
hallucinated: bool = False
judge_score: float = 0.0 # 0β10, from LLM-Judge
gap_signals: list[str] = field(default_factory=list)
@dataclass
class EvalResult:
model_id: str
dataset: str
n_samples: int
avg_rouge1: float = 0.0
avg_rouge2: float = 0.0
avg_rougeL: float = 0.0
avg_bertscore: float = 0.0
avg_faithfulness: float = 0.0
hallucination_rate: float = 0.0
avg_judge_score: float = 0.0
gap_categories: list[str] = field(default_factory=list)
sample_scores: list[SampleScore] = field(default_factory=list)
# βββββββββββββββββββββββββββββββββββββββββββββ
# 1. ROUGE scorer
# βββββββββββββββββββββββββββββββββββββββββββββ
class RougeScorer:
def __init__(self):
from rouge_score import rouge_scorer as rs
self._scorer = rs.RougeScorer(ROUGE_TYPES, use_stemmer=True)
def score(self, prediction: str, reference: str) -> dict[str, float]:
scores = self._scorer.score(reference, prediction)
return {k: scores[k].fmeasure for k in ROUGE_TYPES}
# βββββββββββββββββββββββββββββββββββββββββββββ
# 2. BERTScore
# βββββββββββββββββββββββββββββββββββββββββββββ
class BertScorer:
def __init__(self, lang: str = "en", device: Optional[str] = None):
self.lang = lang
self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")
self._model = None # lazy load
def _load(self):
if self._model is None:
from bert_score import BERTScorer
self._model = BERTScorer(lang=self.lang, device=self.device, rescale_with_baseline=True)
def score_batch(self, predictions: list[str], references: list[str]) -> list[float]:
self._load()
_, _, f1 = self._model.score(predictions, references)
return f1.tolist()
# βββββββββββββββββββββββββββββββββββββββββββββ
# 3. Faithfulness / Hallucination Detector
# βββββββββββββββββββββββββββββββββββββββββββββ
HALLUCINATION_PATTERNS = [
r"\b(definitely|certainly|always|never|everyone|nobody)\b",
r"\b(100%|absolutely|guaranteed)\b",
r"\bin \d{4},? .{0,30} (invented|discovered|created|founded)\b",
]
class FaithfulnessChecker:
"""
Lightweight faithfulness checker using:
1. Heuristic hallucination patterns
2. NLI-based entailment (context β prediction)
"""
def __init__(self):
self._nli = None
def _load_nli(self):
if self._nli is None:
from transformers import pipeline
self._nli = pipeline(
"text-classification",
model="cross-encoder/nli-deberta-v3-small",
device=0 if torch.cuda.is_available() else -1,
)
def _heuristic_hallucination(self, prediction: str) -> bool:
return any(re.search(p, prediction, re.IGNORECASE) for p in HALLUCINATION_PATTERNS)
def check(self, context: str, prediction: str) -> tuple[float, bool]:
"""
Returns (faithfulness_score 0-1, is_hallucinated).
faithfulness_score = NLI entailment probability.
"""
hallucinated = self._heuristic_hallucination(prediction)
try:
self._load_nli()
pair = f"{context[:512]} [SEP] {prediction[:256]}"
result = self._nli(pair, truncation=True, max_length=512)[0]
label = result["label"].upper()
score = result["score"]
faith = score if label == "ENTAILMENT" else (1.0 - score if label == "CONTRADICTION" else 0.5)
if faith < 0.3:
hallucinated = True
except Exception as e:
logger.warning(f"NLI check failed: {e}")
faith = 0.5
return faith, hallucinated
def check_batch(self, contexts: list[str], predictions: list[str]) -> list[tuple[float, bool]]:
return [self.check(c, p) for c, p in zip(contexts, predictions)]
# βββββββββββββββββββββββββββββββββββββββββββββ
# 4. LLM-as-Judge
# βββββββββββββββββββββββββββββββββββββββββββββ
JUDGE_PROMPT = """You are an expert evaluator. Score the following answer on a scale of 0-10.
Question: {question}
Reference Answer: {reference}
Model Answer: {prediction}
Evaluation criteria:
- Accuracy (does it match the reference?)
- Completeness (does it cover all key points?)
- Clarity (is it well-expressed?)
- Faithfulness (no hallucinations?)
Respond ONLY with: SCORE: <integer 0-10>\nREASON: <one sentence>
"""
class LLMJudge:
def __init__(self, judge_model_id: str = JUDGE_MODEL):
self.model_id = judge_model_id
self._pipe = None
def _load(self):
if self._pipe is None:
from transformers import pipeline, BitsAndBytesConfig
bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16)
self._pipe = pipeline(
"text-generation",
model=self.model_id,
model_kwargs={"quantization_config": bnb},
device_map="auto",
token=HF_TOKEN or None,
trust_remote_code=True,
)
def score(self, question: str, reference: str, prediction: str) -> tuple[float, str]:
"""Returns (score 0-10, reason string)."""
try:
self._load()
prompt = JUDGE_PROMPT.format(
question=question, reference=reference, prediction=prediction
)
out = self._pipe(prompt, max_new_tokens=64, do_sample=False)[0]["generated_text"]
# Parse response
score_match = re.search(r"SCORE:\s*(\d+)", out)
reason_match = re.search(r"REASON:\s*(.+)", out)
score = int(score_match.group(1)) if score_match else 5
reason = reason_match.group(1).strip() if reason_match else ""
return min(10.0, max(0.0, float(score))), reason
except Exception as e:
logger.warning(f"Judge scoring failed: {e}")
return 5.0, ""
def compare_models(
self,
question: str, reference: str,
pred_a: str, pred_b: str,
model_a_name: str = "A", model_b_name: str = "B",
) -> dict:
score_a, _ = self.score(question, reference, pred_a)
score_b, _ = self.score(question, reference, pred_b)
winner = model_a_name if score_a >= score_b else model_b_name
return {
"winner": winner,
model_a_name: score_a,
model_b_name: score_b,
"delta": abs(score_a - score_b),
}
# βββββββββββββββββββββββββββββββββββββββββββββ
# 5. Knowledge Gap Detector
# βββββββββββββββββββββββββββββββββββββββββββββ
GAP_THRESHOLDS = {
"rouge": 0.30,
"bertscore": 0.50,
"faithfulness": 0.50,
"judge": 5.0,
}
GAP_CATEGORY_KEYWORDS = {
"factual_recall": ["who", "when", "where", "what year", "which country"],
"multi_step_reasoning": ["therefore", "because", "if", "then", "step"],
"numerical": ["how many", "calculate", "percent", "number", "count"],
"code_generation": ["write a function", "implement", "code", "python", "sql"],
"summarization": ["summarize", "summary", "main points", "briefly"],
}
def detect_gaps(sample_scores: list[SampleScore], samples: list[Sample]) -> list[str]:
"""
Analyze per-sample scores to identify systematic weaknesses.
Returns list of gap category labels.
"""
gap_counts: dict[str, int] = {}
failing: list[int] = []
for i, ss in enumerate(sample_scores):
is_failing = (
ss.rouge1 < GAP_THRESHOLDS["rouge"] or
ss.bertscore_f1 < GAP_THRESHOLDS["bertscore"] or
ss.faithfulness < GAP_THRESHOLDS["faithfulness"] or
ss.judge_score < GAP_THRESHOLDS["judge"] or
ss.hallucinated
)
if is_failing:
failing.append(i)
for idx in failing:
question = (samples[idx].question + " " + samples[idx].context).lower()
for cat, keywords in GAP_CATEGORY_KEYWORDS.items():
if any(kw in question for kw in keywords):
gap_counts[cat] = gap_counts.get(cat, 0) + 1
# Return categories with β₯2 failures
gaps = [cat for cat, cnt in gap_counts.items() if cnt >= 2]
logger.info(f"[GapDetect] Identified gaps: {gaps or ['none']}")
return gaps
# βββββββββββββββββββββββββββββββββββββββββββββ
# 6. Inference helper
# βββββββββββββββββββββββββββββββββββββββββββββ
def run_inference(
model_id: str,
samples: list[Sample],
batch_size: int = 4,
) -> list[Sample]:
"""Fill sample.prediction for each sample using the target model."""
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16)
tok = AutoTokenizer.from_pretrained(model_id, token=HF_TOKEN or None, trust_remote_code=True)
if tok.pad_token is None:
tok.pad_token = tok.eos_token
model = AutoModelForCausalLM.from_pretrained(
model_id,
quantization_config=bnb,
device_map="auto",
token=HF_TOKEN or None,
trust_remote_code=True,
)
model.eval()
for i in range(0, len(samples), batch_size):
batch = samples[i : i + batch_size]
prompts = [
f"Context: {s.context}\nQuestion: {s.question}\nAnswer:"
for s in batch
]
inputs = tok(prompts, return_tensors="pt", padding=True,
truncation=True, max_length=1024).to(model.device)
with torch.no_grad():
out_ids = model.generate(**inputs, max_new_tokens=128, do_sample=False)
for j, sample in enumerate(batch):
new_tokens = out_ids[j][inputs["input_ids"].shape[1]:]
sample.prediction = tok.decode(new_tokens, skip_special_tokens=True).strip()
del model
torch.cuda.empty_cache()
return samples
# βββββββββββββββββββββββββββββββββββββββββββββ
# 7. Master evaluation pipeline
# βββββββββββββββββββββββββββββββββββββββββββββ
def evaluate(
model_id: str,
samples: list[Sample],
dataset_name: str = "custom",
run_judge: bool = True,
save: bool = True,
) -> EvalResult:
"""
Full evaluation pipeline for one model on a dataset.
"""
n = len(samples)
logger.info(f"[Eval] Evaluating '{model_id}' on {n} samples...")
# 1. Generate predictions
samples = run_inference(model_id, samples)
# 2. Initialize scorers
rouge = RougeScorer()
bert = BertScorer()
faith = FaithfulnessChecker()
judge = LLMJudge() if run_judge else None
# 3. Score each sample
predictions = [s.prediction for s in samples]
references = [s.reference for s in samples]
# BERTScore in batch
bs_f1s = bert.score_batch(predictions, references)
sample_scores: list[SampleScore] = []
for i, s in enumerate(track(samples, description="Scoring")):
rscores = rouge.score(s.prediction, s.reference)
faith_sc, hallu = faith.check(s.context, s.prediction)
j_score = 5.0
if judge:
j_score, _ = judge.score(s.question, s.reference, s.prediction)
ss = SampleScore(
sample_id = s.id,
rouge1 = rscores["rouge1"],
rouge2 = rscores["rouge2"],
rougeL = rscores["rougeL"],
bertscore_f1 = bs_f1s[i],
faithfulness = faith_sc,
hallucinated = hallu,
judge_score = j_score,
)
sample_scores.append(ss)
# 4. Aggregate
def avg(lst): return sum(lst) / len(lst) if lst else 0.0
result = EvalResult(
model_id = model_id,
dataset = dataset_name,
n_samples = n,
avg_rouge1 = avg([s.rouge1 for s in sample_scores]),
avg_rouge2 = avg([s.rouge2 for s in sample_scores]),
avg_rougeL = avg([s.rougeL for s in sample_scores]),
avg_bertscore = avg([s.bertscore_f1 for s in sample_scores]),
avg_faithfulness = avg([s.faithfulness for s in sample_scores]),
hallucination_rate = avg([float(s.hallucinated) for s in sample_scores]),
avg_judge_score = avg([s.judge_score for s in sample_scores]),
gap_categories = detect_gaps(sample_scores, samples),
sample_scores = sample_scores,
)
# 5. Display
_print_eval_table(result)
# 6. Save
if save:
out = EVAL_DIR / f"eval_{model_id.replace('/', '_')}_{dataset_name}.json"
with open(out, "w") as f:
json.dump(asdict(result), f, indent=2)
logger.info(f"[Eval] Results saved β {out}")
return result
def compare_models(
model_ids: list[str],
samples: list[Sample],
dataset_name: str = "custom",
) -> list[EvalResult]:
"""Evaluate multiple models and print a side-by-side comparison."""
results = [evaluate(mid, samples, dataset_name) for mid in model_ids]
_print_comparison_table(results)
return results
def _print_eval_table(r: EvalResult) -> None:
table = Table(title=f"Evaluation β {r.model_id}", show_lines=True)
table.add_column("Metric", style="cyan")
table.add_column("Score", justify="right", style="green")
table.add_column("Threshold", justify="right", style="dim")
rows = [
("ROUGE-1", f"{r.avg_rouge1:.3f}", f"{GAP_THRESHOLDS['rouge']:.2f}"),
("ROUGE-2", f"{r.avg_rouge2:.3f}", "β"),
("ROUGE-L", f"{r.avg_rougeL:.3f}", "β"),
("BERTScore F1", f"{r.avg_bertscore:.3f}", f"{GAP_THRESHOLDS['bertscore']:.2f}"),
("Faithfulness", f"{r.avg_faithfulness:.3f}", f"{GAP_THRESHOLDS['faithfulness']:.2f}"),
("Hallucination %", f"{r.hallucination_rate*100:.1f}%", "< 10%"),
("Judge Score", f"{r.avg_judge_score:.1f}/10", f"{GAP_THRESHOLDS['judge']:.0f}/10"),
]
for row in rows:
table.add_row(*row)
console.print(table)
if r.gap_categories:
console.print(f"[yellow]β Gaps detected:[/yellow] {', '.join(r.gap_categories)}")
def _print_comparison_table(results: list[EvalResult]) -> None:
table = Table(title="Model Comparison", show_lines=True)
table.add_column("Metric", style="cyan")
for r in results:
table.add_column(r.model_id.split("/")[-1], justify="right")
metrics = [
("ROUGE-1", lambda r: f"{r.avg_rouge1:.3f}"),
("BERTScore", lambda r: f"{r.avg_bertscore:.3f}"),
("Faithfulness",lambda r: f"{r.avg_faithfulness:.3f}"),
("Halluc. %", lambda r: f"{r.hallucination_rate*100:.1f}%"),
("Judge", lambda r: f"{r.avg_judge_score:.1f}"),
]
for label, fn in metrics:
table.add_row(label, *[fn(r) for r in results])
console.print(table)
# βββββββββββββββββββββββββββββββββββββββββββββ
# Dataset loaders
# βββββββββββββββββββββββββββββββββββββββββββββ
def load_squad(n: int = 100) -> list[Sample]:
"""Load SQuAD v2 validation set as Sample list."""
from datasets import load_dataset
ds = load_dataset("squad_v2", split=f"validation[:{n}]")
samples = []
for i, row in enumerate(ds):
if not row["answers"]["text"]:
continue # skip unanswerable
samples.append(Sample(
id = str(i),
context = row["context"],
question = row["question"],
reference = row["answers"]["text"][0],
))
return samples[:n]
def load_custom(path: str) -> list[Sample]:
"""Load custom JSONL dataset: {id, context, question, reference}"""
with open(path) as f:
return [Sample(**json.loads(line)) for line in f]
# βββββββββββββββββββββββββββββββββββββββββββββ
# CLI
# βββββββββββββββββββββββββββββββββββββββββββββ
@app.command()
def run(
model: str = typer.Argument(..., help="Model ID or local path"),
dataset: str = typer.Option("squad", help="squad | custom"),
data_path: str = typer.Option(None, help="Path to custom JSONL (if dataset=custom)"),
n_samples: int = typer.Option(100, help="Number of samples to evaluate"),
judge: bool = typer.Option(True, "--judge/--no-judge"),
compare: list[str] = typer.Option([], "--compare", help="Additional models for comparison"),
):
if dataset == "squad":
samples = load_squad(n_samples)
elif dataset == "custom" and data_path:
samples = load_custom(data_path)[:n_samples]
else:
raise typer.BadParameter("Provide --data-path for custom dataset")
if compare:
compare_models([model] + list(compare), samples, dataset)
else:
evaluate(model, samples, dataset, run_judge=judge)
if __name__ == "__main__":
app()
|