syscred_duplicate / syscred /run_liar_benchmark.py
D Ф m i И i q ц e L Ф y e r
Deploy SysCRED v2.3.1 - GraphRAG + LIAR benchmark + TREC integration
8e97fc5
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
LIAR Benchmark Runner - SysCRED
================================
Scientific evaluation of SysCRED on the LIAR benchmark dataset.
Usage:
python run_liar_benchmark.py --split test
python run_liar_benchmark.py --sample 100 --verbose
python run_liar_benchmark.py --split test --output results/liar_benchmark.csv
(c) Dominique S. Loyer - PhD Thesis Prototype
"""
import argparse
import json
import time
import sys
from pathlib import Path
from datetime import datetime
from typing import Dict, List, Any, Optional
from collections import Counter
# Add parent to path for imports
sys.path.insert(0, str(Path(__file__).parent.parent))
try:
import pandas as pd
HAS_PANDAS = True
except ImportError:
HAS_PANDAS = False
print("[Warning] pandas not installed. CSV export will be limited.")
try:
from sklearn.metrics import (
accuracy_score, precision_score, recall_score, f1_score,
confusion_matrix, classification_report
)
HAS_SKLEARN = True
except ImportError:
HAS_SKLEARN = False
print("[Warning] sklearn not installed. Using basic metrics.")
from syscred.liar_dataset import LIARDataset, LiarStatement
from syscred.verification_system import CredibilityVerificationSystem
from syscred import config
class LIARBenchmark:
"""
Benchmark runner for evaluating SysCRED on LIAR dataset.
"""
# Map SysCRED score to binary label
SYSCRED_THRESHOLD = 0.5 # Below = Fake, Above = Real
def __init__(
self,
data_dir: Optional[str] = None,
load_ml: bool = True,
use_graphrag: bool = True
):
"""
Initialize benchmark.
Args:
data_dir: Path to LIAR dataset directory
load_ml: Whether to load ML models
use_graphrag: Whether to use GraphRAG context
"""
print("=" * 60)
print("SysCRED LIAR Benchmark Runner")
print("=" * 60)
# Load dataset
self.dataset = LIARDataset(data_dir)
# Initialize SysCRED
print("\n[Benchmark] Initializing SysCRED...")
self.system = CredibilityVerificationSystem(
ontology_base_path=str(config.Config.ONTOLOGY_BASE_PATH),
ontology_data_path=str(config.Config.ONTOLOGY_DATA_PATH),
load_ml_models=load_ml,
google_api_key=config.Config.GOOGLE_FACT_CHECK_API_KEY
)
self.use_graphrag = use_graphrag
self.results: List[Dict] = []
print("[Benchmark] System ready.\n")
def _syscred_to_binary(self, score: float) -> str:
"""Convert SysCRED score to binary label."""
return "Real" if score >= self.SYSCRED_THRESHOLD else "Fake"
def _syscred_to_ternary(self, score: float) -> str:
"""Convert SysCRED score to ternary label."""
if score >= 0.65:
return "True"
elif score >= 0.35:
return "Mixed"
else:
return "False"
def evaluate_statement(self, statement: LiarStatement) -> Dict[str, Any]:
"""
Evaluate a single statement.
Args:
statement: LiarStatement to evaluate
Returns:
Result dictionary with prediction and ground truth
"""
start_time = time.time()
result = {
'id': statement.id,
'statement': statement.statement[:200],
'ground_truth_6way': statement.label.name,
'ground_truth_binary': statement.binary_label,
'ground_truth_ternary': statement.ternary_label,
'speaker': statement.speaker,
'party': statement.party,
'syscred_score': 0.5,
'predicted_binary': 'Unknown',
'predicted_ternary': 'Unknown',
'binary_correct': False,
'ternary_correct': False,
'processing_time': 0,
'error': None
}
try:
# Run SysCRED analysis on the statement text
# Note: LIAR statements are short claims, not URLs
report = self.system.verify_information(statement.statement)
if 'error' not in report:
score = report.get('scoreCredibilite', 0.5)
result['syscred_score'] = score
result['predicted_binary'] = self._syscred_to_binary(score)
result['predicted_ternary'] = self._syscred_to_ternary(score)
# Check correctness
result['binary_correct'] = (result['predicted_binary'] == result['ground_truth_binary'])
result['ternary_correct'] = (result['predicted_ternary'] == result['ground_truth_ternary'])
# Add extra details if available
if 'analyseNLP' in report:
result['sentiment'] = report['analyseNLP'].get('sentiment', {})
result['bias'] = report['analyseNLP'].get('bias_analysis', {})
else:
result['error'] = report['error']
except Exception as e:
result['error'] = str(e)
result['processing_time'] = time.time() - start_time
return result
def run_benchmark(
self,
split: str = "test",
sample_size: Optional[int] = None,
verbose: bool = False
) -> Dict[str, Any]:
"""
Run full benchmark on a dataset split.
Args:
split: 'train', 'valid', or 'test'
sample_size: If set, only evaluate this many statements
verbose: Print progress for each statement
Returns:
Dictionary with metrics and detailed results
"""
print(f"\n[Benchmark] Running on {split} split...")
# Load dataset
statements = self.dataset.load_split(split)
if sample_size:
import random
statements = random.sample(statements, min(sample_size, len(statements)))
print(f"[Benchmark] Using sample of {len(statements)} statements")
total = len(statements)
self.results = []
# Progress tracking
start_time = time.time()
for i, stmt in enumerate(statements):
if verbose or (i + 1) % 50 == 0:
print(f"[{i+1}/{total}] Processing: {stmt.statement[:50]}...")
result = self.evaluate_statement(stmt)
self.results.append(result)
if verbose:
symbol = "✅" if result['binary_correct'] else "❌"
print(f" -> Score: {result['syscred_score']:.2f} | "
f"Pred: {result['predicted_binary']} | "
f"True: {result['ground_truth_binary']} {symbol}")
elapsed = time.time() - start_time
# Calculate metrics
metrics = self._calculate_metrics()
metrics['elapsed_time'] = elapsed
metrics['statements_per_second'] = total / elapsed if elapsed > 0 else 0
return metrics
def _calculate_metrics(self) -> Dict[str, Any]:
"""Calculate evaluation metrics from results."""
if not self.results:
return {'error': 'No results to evaluate'}
# Filter successful evaluations
valid_results = [r for r in self.results if r['error'] is None]
error_count = len(self.results) - len(valid_results)
if not valid_results:
return {'error': 'All evaluations failed'}
metrics = {
'total_statements': len(self.results),
'successful_evaluations': len(valid_results),
'error_count': error_count,
'error_rate': error_count / len(self.results)
}
# Extract labels
y_true_binary = [r['ground_truth_binary'] for r in valid_results]
y_pred_binary = [r['predicted_binary'] for r in valid_results]
y_true_ternary = [r['ground_truth_ternary'] for r in valid_results]
y_pred_ternary = [r['predicted_ternary'] for r in valid_results]
# Binary metrics
if HAS_SKLEARN:
metrics['binary'] = {
'accuracy': accuracy_score(y_true_binary, y_pred_binary),
'precision': precision_score(y_true_binary, y_pred_binary, pos_label='Fake', zero_division=0),
'recall': recall_score(y_true_binary, y_pred_binary, pos_label='Fake', zero_division=0),
'f1': f1_score(y_true_binary, y_pred_binary, pos_label='Fake', zero_division=0),
'confusion_matrix': confusion_matrix(y_true_binary, y_pred_binary, labels=['Fake', 'Real']).tolist()
}
metrics['ternary'] = {
'accuracy': accuracy_score(y_true_ternary, y_pred_ternary),
'macro_f1': f1_score(y_true_ternary, y_pred_ternary, average='macro', zero_division=0),
'confusion_matrix': confusion_matrix(y_true_ternary, y_pred_ternary,
labels=['False', 'Mixed', 'True']).tolist()
}
# Detailed classification report
metrics['classification_report'] = classification_report(
y_true_binary, y_pred_binary,
target_names=['Fake', 'Real'],
output_dict=True
)
else:
# Basic metrics without sklearn
correct_binary = sum(1 for r in valid_results if r['binary_correct'])
correct_ternary = sum(1 for r in valid_results if r['ternary_correct'])
metrics['binary'] = {
'accuracy': correct_binary / len(valid_results),
'correct': correct_binary,
'incorrect': len(valid_results) - correct_binary
}
metrics['ternary'] = {
'accuracy': correct_ternary / len(valid_results),
'correct': correct_ternary,
'incorrect': len(valid_results) - correct_ternary
}
# Score distribution
scores = [r['syscred_score'] for r in valid_results]
metrics['score_distribution'] = {
'mean': sum(scores) / len(scores),
'min': min(scores),
'max': max(scores),
'median': sorted(scores)[len(scores) // 2]
}
# Per-party analysis
party_results = {}
for party in ['republican', 'democrat']:
party_items = [r for r in valid_results if r['party'].lower() == party]
if party_items:
party_correct = sum(1 for r in party_items if r['binary_correct'])
party_results[party] = {
'count': len(party_items),
'accuracy': party_correct / len(party_items)
}
metrics['per_party'] = party_results
return metrics
def print_results(self, metrics: Dict[str, Any]) -> None:
"""Pretty-print benchmark results."""
print("\n" + "=" * 60)
print("LIAR BENCHMARK RESULTS")
print("=" * 60)
print(f"\n📊 Overview:")
print(f" Total Statements: {metrics.get('total_statements', 0)}")
print(f" Successful: {metrics.get('successful_evaluations', 0)}")
print(f" Errors: {metrics.get('error_count', 0)} ({metrics.get('error_rate', 0):.1%})")
print(f" Processing Time: {metrics.get('elapsed_time', 0):.1f}s")
print(f" Speed: {metrics.get('statements_per_second', 0):.2f} stmt/sec")
if 'binary' in metrics:
print(f"\n📈 Binary Classification (Fake vs Real):")
b = metrics['binary']
print(f" Accuracy: {b.get('accuracy', 0):.2%}")
print(f" Precision: {b.get('precision', 0):.2%}")
print(f" Recall: {b.get('recall', 0):.2%}")
print(f" F1-Score: {b.get('f1', 0):.2f}")
if 'confusion_matrix' in b:
cm = b['confusion_matrix']
print(f"\n Confusion Matrix:")
print(f" Pred Fake Pred Real")
print(f" True Fake {cm[0][0]:5d} {cm[0][1]:5d}")
print(f" True Real {cm[1][0]:5d} {cm[1][1]:5d}")
if 'ternary' in metrics:
print(f"\n📊 Ternary Classification (False/Mixed/True):")
t = metrics['ternary']
print(f" Accuracy: {t.get('accuracy', 0):.2%}")
print(f" Macro F1: {t.get('macro_f1', 0):.2f}")
if 'per_party' in metrics:
print(f"\n🏛️ Per-Party Analysis:")
for party, data in metrics['per_party'].items():
print(f" {party.capitalize()}: {data['accuracy']:.2%} accuracy ({data['count']} samples)")
if 'score_distribution' in metrics:
print(f"\n📉 Score Distribution:")
sd = metrics['score_distribution']
print(f" Mean: {sd['mean']:.3f}")
print(f" Median: {sd['median']:.3f}")
print(f" Range: [{sd['min']:.3f}, {sd['max']:.3f}]")
print("\n" + "=" * 60)
def save_results(self, output_path: str, metrics: Dict[str, Any]) -> None:
"""Save results to files."""
output = Path(output_path)
output.parent.mkdir(parents=True, exist_ok=True)
# Save detailed results as CSV
if HAS_PANDAS and self.results:
df = pd.DataFrame(self.results)
csv_path = output.with_suffix('.csv')
df.to_csv(csv_path, index=False)
print(f"[Benchmark] Results saved to: {csv_path}")
# Save metrics as JSON
json_path = output.with_suffix('.json')
with open(json_path, 'w') as f:
json.dump({
'timestamp': datetime.now().isoformat(),
'dataset': 'LIAR',
'metrics': metrics,
'config': {
'threshold': self.SYSCRED_THRESHOLD,
'use_graphrag': self.use_graphrag,
'weights': dict(self.system.weights)
}
}, f, indent=2, default=str)
print(f"[Benchmark] Metrics saved to: {json_path}")
def main():
parser = argparse.ArgumentParser(description='Run LIAR benchmark on SysCRED')
parser.add_argument('--split', type=str, default='test',
choices=['train', 'valid', 'test'],
help='Dataset split to evaluate')
parser.add_argument('--sample', type=int, default=None,
help='Number of statements to sample (for quick testing)')
parser.add_argument('--data-dir', type=str, default=None,
help='Path to LIAR dataset directory')
parser.add_argument('--output', type=str, default=None,
help='Output path for results (CSV/JSON)')
parser.add_argument('--no-ml', action='store_true',
help='Disable ML models for faster testing')
parser.add_argument('--verbose', '-v', action='store_true',
help='Print details for each statement')
args = parser.parse_args()
# Run benchmark
benchmark = LIARBenchmark(
data_dir=args.data_dir,
load_ml=not args.no_ml
)
try:
metrics = benchmark.run_benchmark(
split=args.split,
sample_size=args.sample,
verbose=args.verbose
)
benchmark.print_results(metrics)
if args.output:
benchmark.save_results(args.output, metrics)
else:
# Default output path
default_output = Path(__file__).parent / f"liar_benchmark_{args.split}.csv"
benchmark.save_results(str(default_output), metrics)
except FileNotFoundError as e:
print(f"\n❌ Error: {e}")
print("\nTo download the LIAR dataset:")
print(" 1. wget https://www.cs.ucsb.edu/~william/data/liar_dataset.zip")
print(" 2. unzip liar_dataset.zip -d 02_Code/syscred/datasets/liar/")
sys.exit(1)
if __name__ == "__main__":
main()