File size: 13,973 Bytes
d38c1d3 | 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 | """
Benchmark harness for the reasoning engine.
Train/test split β teach train β evaluate on held-out test.
Reports: exact match, F1, abstention rate, latency.
Usage:
python3 benchmark.py # all datasets, 80/20 split
python3 benchmark.py --split 0.9 # 90/10 split
python3 benchmark.py --dataset hotpotqa # single dataset
python3 benchmark.py --test-only # skip training, test existing brain
"""
import sys, os, json, time, argparse, random, re
from pathlib import Path
from collections import defaultdict
sys.path.insert(0, os.path.join(os.path.dirname(__file__), 'src'))
from brain import Brain
def load_brain(db_path):
"""Load brain β auto-detect CSR > LMDB > SQLite."""
csr_path = os.path.join(db_path, 'cooc_csr', 'indptr.bin')
lmdb_path = os.path.join(db_path, 'brain.lmdb')
if os.path.exists(csr_path) and os.path.exists(lmdb_path):
from brain_csr_adapter import BrainCSR
return BrainCSR(db_path=db_path)
if os.path.exists(lmdb_path):
from brain_lmdb_adapter import BrainLMDB
return BrainLMDB(db_path=db_path)
return Brain(db_path=db_path)
DATA_DIR = Path.home() / "webmind-research" / "data"
def tokenize(text):
"""Simple tokenization for F1 scoring."""
return re.findall(r'\w+', text.lower())
def exact_match(prediction, gold):
"""Normalized exact match."""
return prediction.strip().lower() == gold.strip().lower()
def f1_score(prediction, gold):
"""Token-level F1 between prediction and gold."""
pred_tokens = set(tokenize(prediction))
gold_tokens = set(tokenize(gold))
if not gold_tokens:
return 1.0 if not pred_tokens else 0.0
if not pred_tokens:
return 0.0
common = pred_tokens & gold_tokens
if not common:
return 0.0
precision = len(common) / len(pred_tokens)
recall = len(common) / len(gold_tokens)
return 2 * precision * recall / (precision + recall)
def load_qa_records(dataset_path):
"""Load records that have questions and answers (for evaluation)."""
records = []
with open(dataset_path) as f:
for line in f:
try:
item = json.loads(line.strip())
except json.JSONDecodeError:
continue
question = item.get('question', item.get('text', '')).strip()
answer = item.get('answer', '').strip()
if question and answer:
records.append({'question': question, 'answer': answer})
return records
def split_data(records, train_ratio=0.8, seed=42):
"""Deterministic train/test split."""
rng = random.Random(seed)
shuffled = list(records)
rng.shuffle(shuffled)
split_idx = int(len(shuffled) * train_ratio)
return shuffled[:split_idx], shuffled[split_idx:]
def teach_records(brain, records):
"""Teach training records to brain."""
for rec in records:
q = rec['question']
a = rec['answer']
# Teach the answer as knowledge
if len(a) < 50 and not a.endswith('.'):
combined = f"{q.rstrip('?')} is {a}"
else:
combined = a
brain.teach(combined, confidence=0.6)
def evaluate(brain, test_records, verbose=False):
"""Evaluate brain on held-out test records."""
results = {
'total': len(test_records),
'exact_match': 0,
'f1_sum': 0.0,
'abstentions': 0,
'latency_ms': [],
'examples': [],
}
for rec in test_records:
question = rec['question']
gold = rec['answer']
t0 = time.time()
try:
result = brain.ask(question, auto_learn=False)
except TypeError:
result = brain.ask(question)
latency = (time.time() - t0) * 1000
prediction = result['answer']
strategy = result['strategy']
confidence = result['confidence']
results['latency_ms'].append(latency)
if strategy == 'abstain':
results['abstentions'] += 1
em = 0
f1 = 0.0
else:
em = 1 if exact_match(prediction, gold) else 0
f1 = f1_score(prediction, gold)
results['exact_match'] += em
results['f1_sum'] += f1
if verbose and len(results['examples']) < 10:
results['examples'].append({
'question': question[:80],
'gold': gold[:80],
'prediction': prediction[:80],
'em': em,
'f1': round(f1, 3),
'confidence': round(confidence, 3),
'strategy': strategy,
'latency_ms': round(latency, 1),
})
n = results['total']
results['exact_match_pct'] = round(100 * results['exact_match'] / n, 2) if n else 0
results['f1_avg'] = round(results['f1_sum'] / n, 4) if n else 0
results['abstention_rate'] = round(100 * results['abstentions'] / n, 2) if n else 0
results['latency_avg_ms'] = round(sum(results['latency_ms']) / n, 1) if n else 0
results['latency_p95_ms'] = round(sorted(results['latency_ms'])[int(n * 0.95)] if n else 0, 1)
return results
def rlhf_epoch(brain, test_records, epoch_num):
"""
One RLHF epoch: evaluate β reinforce good edges β weaken bad ones.
For each test question:
- Ask the brain (get answer + participating words)
- Score against gold (F1)
- If F1 > 0.5: boost co-occurrence edges for answer words (Γ1.1)
- If F1 < 0.2: weaken co-occurrence edges (Γ0.9)
- Between: no change (uncertain)
Works with both BrainCSR (WAL) and dict-based Brain (_cooc).
"""
f1_sum = 0.0
em_count = 0
reinforced = 0
weakened = 0
# Check which backend we're using
has_wal = hasattr(brain, '_wal')
has_cooc = hasattr(brain, '_cooc')
for rec in test_records:
question = rec['question']
gold = rec['answer']
try:
result = brain.ask(question, auto_learn=False)
except TypeError:
result = brain.ask(question)
prediction = result['answer']
strategy = result['strategy']
if strategy == 'abstain':
continue
f1 = f1_score(prediction, gold)
f1_sum += f1
if exact_match(prediction, gold):
em_count += 1
# Get word indices that participated in the answer
content_words = [t for t in tokenize(prediction)
if t in brain._word_idx]
word_indices = [brain._word_idx[w] for w in content_words]
if not word_indices:
continue
# Reinforce or weaken based on quality
if f1 > 0.5:
# Good answer β reinforce edges
for widx in word_indices:
profile = brain._get_profile(widx)
for neighbor, weight in profile.items():
if neighbor == widx:
continue
boost = weight * 0.5 # 50% boost (was 10%)
if has_wal:
brain._wal.add_edge(widx, neighbor, boost)
elif has_cooc and widx in brain._cooc:
brain._cooc[widx][neighbor] = weight * 1.5
reinforced += 1
elif f1 < 0.2:
# Bad answer β weaken edges AND teach correct answer via QβA map
if hasattr(brain, 'correct'):
brain.correct(question, gold) # direct QβA mapping + strong edge boost
for widx in word_indices:
profile = brain._get_profile(widx)
for neighbor, weight in profile.items():
if neighbor == widx:
continue
penalty = -weight * 0.5 # 50% reduction (was 10%)
if has_wal:
brain._wal.add_edge(widx, neighbor, penalty)
elif has_cooc and widx in brain._cooc:
brain._cooc[widx][neighbor] = weight * 0.5
weakened += 1
n = len(test_records)
f1_avg = f1_sum / n if n else 0
em_pct = 100 * em_count / n if n else 0
print(f" Epoch {epoch_num}: F1={f1_avg:.4f} EM={em_pct:.1f}% "
f"reinforced={reinforced} weakened={weakened}")
return f1_avg, em_pct, reinforced, weakened
def run_benchmark(args):
"""Main benchmark flow."""
# Find datasets
if args.dataset == 'all':
ds_paths = sorted(DATA_DIR.glob("*.jsonl"))
else:
ds_paths = [DATA_DIR / f"{args.dataset}.jsonl"]
# Load all Q&A records
all_records = []
ds_counts = {}
for ds_path in ds_paths:
if not ds_path.exists() or ds_path.stat().st_size == 0:
continue
records = load_qa_records(ds_path)
if records:
ds_counts[ds_path.stem] = len(records)
all_records.extend(records)
if not all_records:
print("No Q&A records found.")
return
print(f"Loaded {len(all_records):,} Q&A records from {len(ds_counts)} datasets")
for ds, count in sorted(ds_counts.items(), key=lambda x: -x[1])[:10]:
print(f" {ds}: {count:,}")
# Split
train, test = split_data(all_records, train_ratio=args.split)
print(f"\nSplit: {len(train):,} train / {len(test):,} test "
f"({args.split:.0%}/{1-args.split:.0%})")
# Cap test size for speed
if args.test_limit and len(test) > args.test_limit:
test = test[:args.test_limit]
print(f"Test capped at {args.test_limit}")
# Train
if not args.test_only:
db_path = args.db_path or '/tmp/benchmark_brain'
os.makedirs(db_path, exist_ok=True)
print(f"\nTraining on {len(train):,} records...")
brain = Brain(db_path=db_path)
brain.begin_bulk()
t0 = time.time()
teach_records(brain, train)
brain.end_bulk()
train_time = time.time() - t0
print(f"Training: {train_time:.1f}s | "
f"{len(brain._words):,} words | "
f"{brain.db.count():,} neurons")
else:
db_path = args.db_path or os.path.expanduser('~/nexus-brain')
brain = load_brain(db_path)
print(f"\nUsing existing brain: {len(brain._words):,} words")
# RLHF epochs (if requested)
if args.epochs > 0:
print(f"\n=== RLHF: {args.epochs} epochs ===")
# Use a subset for RLHF (faster iterations)
rlhf_set = test[:min(200, len(test))]
epoch_results = []
for epoch in range(1, args.epochs + 1):
f1, em, reinforced, weakened = rlhf_epoch(brain, rlhf_set, epoch)
epoch_results.append({'epoch': epoch, 'f1': f1, 'em': em,
'reinforced': reinforced, 'weakened': weakened})
if reinforced == 0 and weakened == 0:
print(f" Converged at epoch {epoch} (no changes)")
break
print(f" RLHF complete. F1 improvement: "
f"{epoch_results[0]['f1']:.4f} β {epoch_results[-1]['f1']:.4f}")
# Evaluate (final, after RLHF if any)
print(f"\nEvaluating on {len(test):,} held-out questions...")
t0 = time.time()
results = evaluate(brain, test, verbose=True)
eval_time = time.time() - t0
# Report
print(f"\n{'='*60}")
print(f"BENCHMARK RESULTS")
print(f"{'='*60}")
print(f" Exact Match: {results['exact_match_pct']}% "
f"({results['exact_match']}/{results['total']})")
print(f" Token F1: {results['f1_avg']}")
print(f" Abstention Rate: {results['abstention_rate']}% "
f"({results['abstentions']}/{results['total']})")
print(f" Latency (avg): {results['latency_avg_ms']} ms")
print(f" Latency (p95): {results['latency_p95_ms']} ms")
print(f" Eval time: {eval_time:.1f}s")
print(f"{'='*60}")
if results['examples']:
print(f"\nSample predictions:")
for ex in results['examples'][:5]:
status = "β" if ex['em'] else ("β
" if ex['strategy'] == 'abstain' else "β")
print(f" {status} Q: {ex['question']}")
print(f" Gold: {ex['gold']}")
print(f" Pred: {ex['prediction']} "
f"[{ex['strategy']}, f1={ex['f1']}, {ex['latency_ms']}ms]")
print()
# Save results
out_path = Path(db_path) / 'benchmark_results.json'
with open(out_path, 'w') as f:
json.dump({
'exact_match_pct': results['exact_match_pct'],
'f1_avg': results['f1_avg'],
'abstention_rate': results['abstention_rate'],
'latency_avg_ms': results['latency_avg_ms'],
'latency_p95_ms': results['latency_p95_ms'],
'total_test': results['total'],
'total_train': len(train),
'datasets': ds_counts,
'split': args.split,
}, f, indent=2)
print(f"Results saved to {out_path}")
brain.close()
def main():
parser = argparse.ArgumentParser(description="Benchmark the reasoning engine")
parser.add_argument("--dataset", default="all",
help="Dataset name or 'all'")
parser.add_argument("--split", type=float, default=0.8,
help="Train ratio (default 0.8 = 80/20)")
parser.add_argument("--test-limit", type=int, default=500,
help="Max test questions for speed (0=unlimited)")
parser.add_argument("--test-only", action="store_true",
help="Skip training, evaluate existing brain")
parser.add_argument("--db-path", default=None,
help="Brain DB path (default: /tmp/benchmark_brain)")
parser.add_argument("--epochs", type=int, default=5,
help="RLHF epochs (0=skip, default=5)")
parser.add_argument("--seed", type=int, default=42)
args = parser.parse_args()
run_benchmark(args)
if __name__ == '__main__':
main()
|