Spaces:
Sleeping
Sleeping
File size: 10,487 Bytes
3c1c6ff |
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 |
#!/usr/bin/env python3
"""
Load Questions from HuggingFace Big Benchmarks Collection
==========================================================
Loads benchmark questions from multiple sources to achieve 20+ domain coverage:
1. MMLU - 57 subjects (already have 14K)
2. ARC-Challenge - Science reasoning
3. HellaSwag - Commonsense NLI
4. TruthfulQA - Truthfulness detection
5. GSM8K - Math word problems
6. Winogrande - Commonsense reasoning
7. BBH - Big-Bench Hard (23 challenging tasks)
Target: 20+ domains with 20,000+ total questions
"""
from pathlib import Path
from benchmark_vector_db import BenchmarkVectorDB, BenchmarkQuestion
from datasets import load_dataset
import logging
from typing import List
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)
def load_arc_challenge() -> List[BenchmarkQuestion]:
"""
Load ARC-Challenge - Science reasoning questions
Domain: Science (physics, chemistry, biology)
Difficulty: Moderate-Hard (GPT-3 ~50%)
"""
logger.info("Loading ARC-Challenge dataset...")
questions = []
try:
dataset = load_dataset("allenai/ai2_arc", "ARC-Challenge", split="test")
logger.info(f" Loaded {len(dataset)} ARC-Challenge questions")
for idx, item in enumerate(dataset):
question = BenchmarkQuestion(
question_id=f"arc_challenge_{idx}",
source_benchmark="ARC-Challenge",
domain="science",
question_text=item['question'],
correct_answer=item['answerKey'],
choices=item['choices']['text'] if 'choices' in item else [],
success_rate=0.50, # Moderate difficulty
difficulty_score=0.50,
difficulty_label="Moderate",
num_models_tested=0
)
questions.append(question)
logger.info(f" ✓ Loaded {len(questions)} science reasoning questions")
except Exception as e:
logger.error(f"Failed to load ARC-Challenge: {e}")
return questions
def load_hellaswag() -> List[BenchmarkQuestion]:
"""
Load HellaSwag - Commonsense NLI
Domain: Commonsense reasoning
Difficulty: Moderate (GPT-3 ~78%)
"""
logger.info("Loading HellaSwag dataset...")
questions = []
try:
dataset = load_dataset("Rowan/hellaswag", split="validation")
logger.info(f" Loaded {len(dataset)} HellaSwag questions")
# Sample to manage size (10K is huge)
max_samples = 2000
if len(dataset) > max_samples:
import random
indices = random.sample(range(len(dataset)), max_samples)
dataset = dataset.select(indices)
for idx, item in enumerate(dataset):
question = BenchmarkQuestion(
question_id=f"hellaswag_{idx}",
source_benchmark="HellaSwag",
domain="commonsense",
question_text=item['ctx'],
correct_answer=str(item['label']),
choices=item['endings'] if 'endings' in item else [],
success_rate=0.65, # Moderate difficulty
difficulty_score=0.35,
difficulty_label="Moderate",
num_models_tested=0
)
questions.append(question)
logger.info(f" ✓ Loaded {len(questions)} commonsense reasoning questions")
except Exception as e:
logger.error(f"Failed to load HellaSwag: {e}")
return questions
def load_gsm8k() -> List[BenchmarkQuestion]:
"""
Load GSM8K - Math word problems
Domain: Mathematics (grade school word problems)
Difficulty: Moderate-Hard (GPT-3 ~35%, GPT-4 ~92%)
"""
logger.info("Loading GSM8K dataset...")
questions = []
try:
dataset = load_dataset("openai/gsm8k", "main", split="test")
logger.info(f" Loaded {len(dataset)} GSM8K questions")
for idx, item in enumerate(dataset):
question = BenchmarkQuestion(
question_id=f"gsm8k_{idx}",
source_benchmark="GSM8K",
domain="math_word_problems",
question_text=item['question'],
correct_answer=item['answer'],
choices=None, # Free-form answer
success_rate=0.55, # Moderate-Hard
difficulty_score=0.45,
difficulty_label="Moderate",
num_models_tested=0
)
questions.append(question)
logger.info(f" ✓ Loaded {len(questions)} math word problem questions")
except Exception as e:
logger.error(f"Failed to load GSM8K: {e}")
return questions
def load_truthfulqa() -> List[BenchmarkQuestion]:
"""
Load TruthfulQA - Truthfulness evaluation
Domain: Truthfulness, factuality
Difficulty: Hard (GPT-3 ~20%, models often confidently wrong)
"""
logger.info("Loading TruthfulQA dataset...")
questions = []
try:
dataset = load_dataset("truthful_qa", "generation", split="validation")
logger.info(f" Loaded {len(dataset)} TruthfulQA questions")
for idx, item in enumerate(dataset):
question = BenchmarkQuestion(
question_id=f"truthfulqa_{idx}",
source_benchmark="TruthfulQA",
domain="truthfulness",
question_text=item['question'],
correct_answer=item['best_answer'],
choices=None,
success_rate=0.35, # Hard - models struggle with truthfulness
difficulty_score=0.65,
difficulty_label="Hard",
num_models_tested=0
)
questions.append(question)
logger.info(f" ✓ Loaded {len(questions)} truthfulness questions")
except Exception as e:
logger.error(f"Failed to load TruthfulQA: {e}")
return questions
def load_winogrande() -> List[BenchmarkQuestion]:
"""
Load Winogrande - Commonsense reasoning
Domain: Commonsense (pronoun resolution)
Difficulty: Moderate (GPT-3 ~70%)
"""
logger.info("Loading Winogrande dataset...")
questions = []
try:
dataset = load_dataset("winogrande", "winogrande_xl", split="validation")
logger.info(f" Loaded {len(dataset)} Winogrande questions")
for idx, item in enumerate(dataset):
question = BenchmarkQuestion(
question_id=f"winogrande_{idx}",
source_benchmark="Winogrande",
domain="commonsense_reasoning",
question_text=item['sentence'],
correct_answer=item['answer'],
choices=[item['option1'], item['option2']],
success_rate=0.70, # Moderate
difficulty_score=0.30,
difficulty_label="Moderate",
num_models_tested=0
)
questions.append(question)
logger.info(f" ✓ Loaded {len(questions)} commonsense reasoning questions")
except Exception as e:
logger.error(f"Failed to load Winogrande: {e}")
return questions
def build_comprehensive_database():
"""Build database with questions from Big Benchmarks Collection"""
logger.info("=" * 70)
logger.info("Loading Questions from Big Benchmarks Collection")
logger.info("=" * 70)
# Initialize database
db = BenchmarkVectorDB(
db_path=Path("./data/benchmark_vector_db"),
embedding_model="all-MiniLM-L6-v2"
)
logger.info(f"\nCurrent database: {db.collection.count():,} questions")
# Load new benchmark datasets
all_new_questions = []
logger.info("\n" + "=" * 70)
logger.info("Phase 1: Science Reasoning (ARC-Challenge)")
logger.info("=" * 70)
arc_questions = load_arc_challenge()
all_new_questions.extend(arc_questions)
logger.info("\n" + "=" * 70)
logger.info("Phase 2: Commonsense NLI (HellaSwag)")
logger.info("=" * 70)
hellaswag_questions = load_hellaswag()
all_new_questions.extend(hellaswag_questions)
logger.info("\n" + "=" * 70)
logger.info("Phase 3: Math Word Problems (GSM8K)")
logger.info("=" * 70)
gsm8k_questions = load_gsm8k()
all_new_questions.extend(gsm8k_questions)
logger.info("\n" + "=" * 70)
logger.info("Phase 4: Truthfulness (TruthfulQA)")
logger.info("=" * 70)
truthfulqa_questions = load_truthfulqa()
all_new_questions.extend(truthfulqa_questions)
logger.info("\n" + "=" * 70)
logger.info("Phase 5: Commonsense Reasoning (Winogrande)")
logger.info("=" * 70)
winogrande_questions = load_winogrande()
all_new_questions.extend(winogrande_questions)
# Index all new questions
logger.info("\n" + "=" * 70)
logger.info(f"Indexing {len(all_new_questions):,} NEW questions")
logger.info("=" * 70)
if all_new_questions:
db.index_questions(all_new_questions)
# Final stats
final_count = db.collection.count()
logger.info("\n" + "=" * 70)
logger.info("FINAL DATABASE STATISTICS")
logger.info("=" * 70)
logger.info(f"\nTotal Questions: {final_count:,}")
logger.info(f"New Questions Added: {len(all_new_questions):,}")
logger.info(f"Previous Count: {final_count - len(all_new_questions):,}")
# Get domain breakdown
sample = db.collection.get(limit=min(5000, final_count), include=['metadatas'])
domains = {}
for meta in sample['metadatas']:
domain = meta.get('domain', 'unknown')
domains[domain] = domains.get(domain, 0) + 1
logger.info(f"\nDomains Found (from sample of {len(sample['metadatas'])}): {len(domains)}")
for domain, count in sorted(domains.items(), key=lambda x: x[1], reverse=True):
logger.info(f" {domain:30} {count:5} questions")
logger.info("\n" + "=" * 70)
logger.info("✅ Database expansion complete!")
logger.info("=" * 70)
return db
if __name__ == "__main__":
build_comprehensive_database()
logger.info("\n🎉 All done! Your database now has comprehensive domain coverage!")
logger.info(" Ready for your VC pitch with 20+ domains! 🚀")
|