Spaces:
Running
Running
File size: 11,063 Bytes
c262d1a |
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 |
#!/usr/bin/env python3
"""
Clean Batch Test - No overrides, pure LLM reasoning with tools
Based on test_specific_question.py but for all questions at once
"""
import os
import sys
import json
import time
from pathlib import Path
from dotenv import load_dotenv
from concurrent.futures import ThreadPoolExecutor, as_completed
# Load environment variables
load_dotenv()
# Add parent directory to path for imports
sys.path.append(str(Path(__file__).parent.parent))
# Local imports
from gaia_web_loader import GAIAQuestionLoaderWeb
from main import GAIASolver
from question_classifier import QuestionClassifier
def load_validation_answers():
"""Load correct answers from GAIA validation metadata"""
answers = {}
try:
validation_path = Path(__file__).parent.parent / 'gaia_validation_metadata.jsonl'
with open(validation_path, 'r') as f:
for line in f:
if line.strip():
data = json.loads(line.strip())
task_id = data.get('task_id')
final_answer = data.get('Final answer')
if task_id and final_answer:
answers[task_id] = final_answer
except Exception as e:
print(f"β οΈ Could not load validation data: {e}")
return answers
def validate_answer(task_id: str, our_answer: str, validation_answers: dict):
"""Validate our answer against the correct answer"""
if task_id not in validation_answers:
return None
expected = str(validation_answers[task_id]).strip()
our_clean = str(our_answer).strip()
# Exact match
if our_clean.lower() == expected.lower():
return {"status": "CORRECT", "expected": expected, "our": our_clean}
# Check if our answer contains the expected answer
if expected.lower() in our_clean.lower():
return {"status": "PARTIAL", "expected": expected, "our": our_clean}
return {"status": "INCORRECT", "expected": expected, "our": our_clean}
def test_single_question(question_data, validation_answers, model="qwen3-235b"):
"""Test a single question without any overrides"""
task_id = question_data.get('task_id', 'unknown')
try:
print(f"π§ͺ [{task_id[:8]}...] Starting...")
# Initialize solver and classifier
solver = GAIASolver(use_kluster=True, kluster_model=model)
classifier = QuestionClassifier()
# Classify the question
question_text = question_data.get('question', '')
file_name = question_data.get('file_name', '')
classification = classifier.classify_question(question_text, file_name)
# Solve the question (NO OVERRIDES - pure LLM reasoning)
start_time = time.time()
answer = solver.solve_question(question_data)
end_time = time.time()
duration = end_time - start_time
# Validate answer
validation_result = validate_answer(task_id, answer, validation_answers)
result = {
'task_id': task_id,
'question_type': classification['primary_agent'],
'complexity': classification['complexity'],
'confidence': classification['confidence'],
'our_answer': str(answer),
'expected_answer': validation_result['expected'] if validation_result else 'N/A',
'status': validation_result['status'] if validation_result else 'NO_VALIDATION',
'duration': duration,
'question_preview': question_data.get('question', '')[:50] + "..."
}
status_icon = "β
" if result['status'] == "CORRECT" else "π‘" if result['status'] == "PARTIAL" else "β"
print(f"{status_icon} [{task_id[:8]}...] {result['status']} | {result['question_type']} | {duration:.1f}s")
return result
except Exception as e:
print(f"β [{task_id[:8]}...] ERROR: {str(e)}")
return {
'task_id': task_id,
'question_type': 'error',
'complexity': 0,
'confidence': 0.0,
'our_answer': '',
'expected_answer': validation_answers.get(task_id, 'N/A'),
'status': 'ERROR',
'duration': 0.0,
'error': str(e),
'question_preview': question_data.get('question', '')[:50] + "..."
}
def run_clean_batch_test():
"""Run clean batch test on all questions"""
print("π§ͺ CLEAN BATCH TEST - NO OVERRIDES")
print("=" * 60)
print("π― Goal: Measure real accuracy with pure LLM reasoning")
print("π« No hardcoded answers or overrides")
print("π€ Pure LLM + Tools reasoning only")
print()
# Load questions and validation data
print("π Loading GAIA questions...")
loader = GAIAQuestionLoaderWeb()
all_questions = loader.questions
validation_answers = load_validation_answers()
print(f"β
Loaded {len(all_questions)} questions")
print(f"β
Loaded {len(validation_answers)} validation answers")
# Show question preview
print(f"\nπ Questions to test:")
for i, q in enumerate(all_questions[:5]): # Show first 5
task_id = q.get('task_id', 'unknown')
question_preview = q.get('question', '')[:40] + "..."
level = q.get('Level', 'Unknown')
has_file = "π" if q.get('file_name') else "π"
print(f" {i+1}. {task_id[:8]}... | L{level} | {has_file} | {question_preview}")
if len(all_questions) > 5:
print(f" ... and {len(all_questions) - 5} more questions")
print(f"\nπ Starting clean batch test...")
print(f"β±οΈ Estimated time: ~{len(all_questions) * 2} minutes")
# Process all questions sequentially (to avoid resource conflicts)
start_time = time.time()
results = []
for i, question_data in enumerate(all_questions):
print(f"\nπ Progress: {i+1}/{len(all_questions)}")
result = test_single_question(question_data, validation_answers)
results.append(result)
end_time = time.time()
total_duration = end_time - start_time
# Analyze results
print(f"\n" + "=" * 60)
print(f"π CLEAN BATCH TEST RESULTS")
print(f"=" * 60)
# Calculate metrics
total_questions = len(results)
correct_answers = len([r for r in results if r['status'] == 'CORRECT'])
partial_answers = len([r for r in results if r['status'] == 'PARTIAL'])
incorrect_answers = len([r for r in results if r['status'] == 'INCORRECT'])
errors = len([r for r in results if r['status'] == 'ERROR'])
accuracy_rate = correct_answers / total_questions * 100
success_rate = (correct_answers + partial_answers) / total_questions * 100
print(f"β±οΈ Total Duration: {int(total_duration // 60)}m {int(total_duration % 60)}s")
print(f"β
Pure Accuracy: {accuracy_rate:.1f}% ({correct_answers}/{total_questions})")
print(f"π― Success Rate: {success_rate:.1f}% (including partial)")
print(f"β‘ Avg per Question: {total_duration/total_questions:.1f}s")
print(f"\nπ DETAILED BREAKDOWN:")
print(f" β
CORRECT: {correct_answers} ({correct_answers/total_questions:.1%})")
print(f" π‘ PARTIAL: {partial_answers} ({partial_answers/total_questions:.1%})")
print(f" β INCORRECT: {incorrect_answers} ({incorrect_answers/total_questions:.1%})")
print(f" π₯ ERROR: {errors} ({errors/total_questions:.1%})")
# Classification performance
print(f"\nπ― CLASSIFICATION PERFORMANCE:")
classification_stats = {}
for result in results:
classification = result['question_type']
if classification not in classification_stats:
classification_stats[classification] = {'total': 0, 'correct': 0, 'partial': 0}
classification_stats[classification]['total'] += 1
if result['status'] == 'CORRECT':
classification_stats[classification]['correct'] += 1
elif result['status'] == 'PARTIAL':
classification_stats[classification]['partial'] += 1
for classification, stats in sorted(classification_stats.items()):
total = stats['total']
correct = stats['correct']
partial = stats['partial']
accuracy = correct / total * 100 if total > 0 else 0
success = (correct + partial) / total * 100 if total > 0 else 0
print(f" {classification:15} | {accuracy:5.1f}% acc | {success:5.1f}% success | {total:2d} questions")
# Detailed results
print(f"\nπ DETAILED QUESTION RESULTS:")
for i, result in enumerate(results):
status_icon = "β
" if result['status'] == "CORRECT" else "π‘" if result['status'] == "PARTIAL" else "β"
print(f" {i+1:2d}. {status_icon} {result['task_id'][:8]}... | {result['question_type']:12} | {result['status']:9} | {result['duration']:5.1f}s")
print(f" Expected: {result['expected_answer']}")
print(f" Got: {result['our_answer']}")
if 'error' in result:
print(f" Error: {result['error']}")
print()
# Save results
timestamp = time.strftime("%Y%m%d_%H%M%S")
results_file = f"logs/clean_batch_test_{timestamp}.json"
with open(results_file, 'w') as f:
json.dump({
'test_metadata': {
'timestamp': timestamp,
'test_type': 'clean_batch_no_overrides',
'total_questions': total_questions,
'duration_seconds': total_duration,
'model': 'qwen3-235b'
},
'metrics': {
'accuracy_rate': accuracy_rate,
'success_rate': success_rate,
'correct_answers': correct_answers,
'partial_answers': partial_answers,
'incorrect_answers': incorrect_answers,
'errors': errors
},
'classification_performance': classification_stats,
'detailed_results': results
}, f, indent=2)
print(f"π Results saved to: {results_file}")
# Final assessment
print(f"\nπ― FINAL ASSESSMENT:")
if accuracy_rate >= 70:
print(f"π EXCELLENT: {accuracy_rate:.1f}% accuracy achieves 70%+ target!")
elif accuracy_rate >= 50:
print(f"π§ GOOD PROGRESS: {accuracy_rate:.1f}% accuracy, approaching target")
elif accuracy_rate >= 30:
print(f"β οΈ MODERATE: {accuracy_rate:.1f}% accuracy, significant room for improvement")
else:
print(f"π¨ NEEDS WORK: {accuracy_rate:.1f}% accuracy requires major improvements")
print(f"\nπ This is the REAL accuracy without any hardcoded answers!")
print(f"π Pure LLM + Tools Performance: {accuracy_rate:.1f}%")
return accuracy_rate, results
if __name__ == "__main__":
accuracy, results = run_clean_batch_test()
print(f"\nπ Clean batch test completed!")
print(f"π Real Accuracy: {accuracy:.1f}%") |