modify files

31 Indobert-base-p1_without_law_with_svr.py

@@ -7,6 +7,7 @@
 import os
 import re
 import json
+import sys
 import zipfile
 import shutil
 import tempfile
@@ -258,9 +259,18 @@ tokenizer.save_pretrained(OUTPUT_DIR)
 
 
 # ======================================================
-# 6. Evaluation
+# 6. Evaluation (Comprehensive)
 # ======================================================
-print("\n📊 Evaluating on Test Set...")
+from sklearn.metrics import (
+    mean_squared_error, 
+    mean_absolute_error, 
+    r2_score,
+    mean_absolute_percentage_error
+)
+
+print("\n" + "="*70)
+print("📊 COMPUTING COMPREHENSIVE TEST SET METRICS")
+print("="*70)
 
 # Predict (Output is scaled)
 preds_scaled = svr_pipeline.predict(X_test)
@@ -272,50 +282,268 @@ y_true_days = label_scaler.inverse_transform(y_test_scaled.reshape(-1, 1)).flatt
 # Clip negatives (sentences can't be negative)
 preds_days = np.maximum(preds_days, 0)
 
-# Metrics
+# Basic metrics
 mse = mean_squared_error(y_true_days, preds_days)
 rmse = np.sqrt(mse)
 mae = mean_absolute_error(y_true_days, preds_days)
 r2 = r2_score(y_true_days, preds_days)
 
-print("\n🎯 Test Results:")
-print(f"  RMSE: {rmse:,.2f} days")
-print(f"  MAE:  {mae:,.2f} days")
-print(f"  R2:   {r2:.4f}")
+# Additional metrics
+median_ae = np.median(np.abs(y_true_days - preds_days))
+max_error = np.max(np.abs(y_true_days - preds_days))
+min_error = np.min(np.abs(y_true_days - preds_days))
+
+# Percentage errors (avoid division by zero)
+mask = y_true_days != 0
+if mask.sum() > 0:
+    mape = np.mean(np.abs((y_true_days[mask] - preds_days[mask]) / y_true_days[mask])) * 100
+else:
+    mape = float('nan')
+
+# Statistics
+pred_mean = np.mean(preds_days)
+pred_std = np.std(preds_days)
+pred_min = np.min(preds_days)
+pred_max = np.max(preds_days)
+
+true_mean = np.mean(y_true_days)
+true_std = np.std(y_true_days)
+true_min = np.min(y_true_days)
+true_max = np.max(y_true_days)
+
+print("\n🎯 ERROR METRICS:")
+print(f"  MSE:                {mse:>12,.2f} days²")
+print(f"  RMSE:               {rmse:>12,.2f} days")
+print(f"  MAE:                {mae:>12,.2f} days")
+print(f"  Median AE:          {median_ae:>12,.2f} days")
+print(f"  Max Error:          {max_error:>12,.2f} days")
+print(f"  Min Error:          {min_error:>12,.2f} days")
+if not np.isnan(mape):
+    print(f"  MAPE:               {mape:>12,.2f}%")
+print(f"  R² Score:           {r2:>12,.4f}")
+
+print("\n📊 PREDICTION STATISTICS:")
+print(f"  Mean:               {pred_mean:>12,.2f} days")
+print(f"  Std Dev:            {pred_std:>12,.2f} days")
+print(f"  Min:                {pred_min:>12,.2f} days")
+print(f"  Max:                {pred_max:>12,.2f} days")
+
+print("\n📊 TRUE LABEL STATISTICS:")
+print(f"  Mean:               {true_mean:>12,.2f} days")
+print(f"  Std Dev:            {true_std:>12,.2f} days")
+print(f"  Min:                {true_min:>12,.2f} days")
+print(f"  Max:                {true_max:>12,.2f} days")
+
+# Error distribution analysis
+print("\n📉 ERROR DISTRIBUTION:")
+errors = np.abs(y_true_days - preds_days)
+percentiles = [10, 25, 50, 75, 90, 95, 99]
+print("  Percentiles of Absolute Error:")
+for p in percentiles:
+    val = np.percentile(errors, p)
+    print(f"    {p}th percentile:   {val:>12,.2f} days")
+
+# Accuracy within ranges
+print("\n🎯 ACCURACY WITHIN ERROR RANGES:")
+for threshold in [100, 250, 500, 750, 1000]:
+    within = np.sum(errors <= threshold) / len(errors) * 100
+    print(f"  Within ±{threshold:>4} days:  {within:>6.2f}% of samples")
 
-# Save detailed results
+
+# ======================================================
+# SAVE DETAILED RESULTS
+# ======================================================
+detailed_results = {
+    'error_metrics': {
+        'mse': float(mse),
+        'rmse': float(rmse),
+        'mae': float(mae),
+        'median_ae': float(median_ae),
+        'max_error': float(max_error),
+        'min_error': float(min_error),
+        'mape': float(mape) if not np.isnan(mape) else None,
+        'r2_score': float(r2)
+    },
+    'prediction_stats': {
+        'mean': float(pred_mean),
+        'std': float(pred_std),
+        'min': float(pred_min),
+        'max': float(pred_max)
+    },
+    'true_label_stats': {
+        'mean': float(true_mean),
+        'std': float(true_std),
+        'min': float(true_min),
+        'max': float(true_max)
+    },
+    'error_percentiles': {
+        f'p{p}': float(np.percentile(errors, p)) for p in percentiles
+    },
+    'accuracy_within_ranges': {
+        f'within_{threshold}': float(np.sum(errors <= threshold) / len(errors) * 100)
+        for threshold in [100, 250, 500, 750, 1000]
+    }
+}
+
+results_detailed_file = Path(OUTPUT_DIR) / 'test_results_detailed.json'
+with open(results_detailed_file, 'w') as f:
+    json.dump(detailed_results, f, indent=2)
+
+print(f"\n💾 Detailed results saved to: {results_detailed_file}")
+
+# Save predictions vs true labels CSV
 results_df = pd.DataFrame({
     'file_name': test_filenames,
     'true_label': y_true_days,
     'prediction': preds_days,
-    'error': np.abs(y_true_days - preds_days)
+    'absolute_error': errors,
+    'percentage_error': np.abs((y_true_days - preds_days) / (y_true_days + 1)) * 100  # +1 to avoid div by 0
 })
-results_csv = Path(OUTPUT_DIR) / 'svr_test_predictions.csv'
+results_csv = Path(OUTPUT_DIR) / 'test_predictions.csv'
 results_df.to_csv(results_csv, index=False)
-print(f"💾 Detailed predictions saved to {results_csv}")
+print(f"💾 Predictions CSV saved to: {results_csv}")
+
+# Also save a simple inference CSV with filename and predicted value (user-friendly export)
+try:
+    # Ensure alignment in case of any mismatch
+    n_simple = min(len(test_filenames), len(preds_days))
+    simple_df = pd.DataFrame({
+        'filename': test_filenames[:n_simple],
+        'predicted_day_sentece': np.array(preds_days[:n_simple]).flatten()
+    })
+    simple_results_file = Path(OUTPUT_DIR) / 'test_inference_simple.csv'
+    simple_df.to_csv(simple_results_file, index=False)
+    print(f"💾 Simple inference CSV saved to: {simple_results_file}")
+except Exception as e:
+    print(f"⚠️ Failed to save simple inference CSV: {e}")
+
+# ======================================================
+# 🧪 SAMPLE PREDICTIONS DISPLAY
+# ======================================================
+print("\n" + "="*70)
+print("🧪 SAMPLE PREDICTIONS (First 10)")
+print("="*70)
+for i in range(min(10, len(preds_days))):
+    fname = test_filenames[i] if i < len(test_filenames) else ''
+    true_val = y_true_days[i]
+    pred_val = preds_days[i]
+    error = abs(true_val - pred_val)
+    pct_error = (error / (true_val + 1)) * 100
+    
+    print(f"\nSample {i+1}:")
+    print(f"  File:       {fname}")
+    print(f"  True:       {true_val:>8,.0f} days")
+    print(f"  Predicted:  {pred_val:>8,.0f} days")
+    print(f"  Error:      {error:>8,.0f} days ({pct_error:.1f}%)")
+
+print("\n" + "="*70)
+print("✅ COMPREHENSIVE TEST SET EVALUATION COMPLETE!")
+print(f"   - {results_detailed_file}")
+print(f"   - {results_csv}")
+print(f"   - {OUTPUT_DIR}/label_scaler.pkl")
+print(f"   - {OUTPUT_DIR}/svr_model.pkl")
 
 
 # ======================================================
 # 7. Plotting
 # ======================================================
-print("\n📈 Generating Plots...")
-img_dir = Path(OUTPUT_DIR) / 'plots'
-img_dir.mkdir(parents=True, exist_ok=True)
-
-# Scatter Plot
-plt.figure(figsize=(8, 8))
-plt.scatter(results_df['true_label'], results_df['prediction'], alpha=0.5, color='purple')
-lims = [0, max(results_df['true_label'].max(), results_df['prediction'].max()) * 1.05]
-plt.plot(lims, lims, '--', color='gray')
-plt.xlim(lims)
-plt.ylim(lims)
-plt.title(f'SVR: True vs Predicted Days (R2={r2:.3f})')
-plt.xlabel('True Days')
-plt.ylabel('Predicted Days')
-plt.grid(True, alpha=0.3)
-plt.tight_layout()
-plt.savefig(img_dir / 'svr_scatter.png')
-print(f"🖼️ Plot saved to {img_dir / 'svr_scatter.png'}")
+try:
+    print("\n🔍 Generating plots for predictions...")
+    img_dir = Path(OUTPUT_DIR) / 'plots'
+    img_dir.mkdir(parents=True, exist_ok=True)
+
+    # Basic scatterplots over sample index
+    idx = np.arange(len(results_df))
+
+    plt.figure(figsize=(12, 4))
+    plt.scatter(idx, results_df['true_label'], s=6, alpha=0.6)
+    plt.title('Ground truth (true_label) over samples')
+    plt.xlabel('Sample index')
+    plt.ylabel('Days')
+    plt.tight_layout()
+    p1 = img_dir / 'scatter_true.png'
+    plt.savefig(p1)
+    plt.close()
+
+    plt.figure(figsize=(12, 4))
+    plt.scatter(idx, results_df['prediction'], s=6, alpha=0.6, color='orange')
+    plt.title('Predictions over samples')
+    plt.xlabel('Sample index')
+    plt.ylabel('Predicted days')
+    plt.tight_layout()
+    p2 = img_dir / 'scatter_pred.png'
+    plt.savefig(p2)
+    plt.close()
+
+    # True vs Predicted scatter (combination)
+    plt.figure(figsize=(6, 6))
+    plt.scatter(results_df['true_label'], results_df['prediction'], s=8, alpha=0.5)
+    lims = [0, max(results_df['true_label'].max(), results_df['prediction'].max()) * 1.05]
+    plt.plot(lims, lims, '--', color='gray')
+    plt.xlim(lims)
+    plt.ylim(lims)
+    plt.xlabel('True days')
+    plt.ylabel('Predicted days')
+    plt.title('True vs Predicted')
+    plt.tight_layout()
+    p3 = img_dir / 'scatter_true_vs_pred.png'
+    plt.savefig(p3)
+    plt.close()
+
+    print(f"✅ Scatter plots saved: {p1}, {p2}, {p3}")
+
+    # Merge with test metadata to get crime_category (charge classification)
+    try:
+        meta_test = pd.read_csv(splits_dir / 'test_split.csv')
+        merged = results_df.merge(meta_test[['file_name', 'crime_category']], on='file_name', how='left')
+
+        # Choose top categories by sample count for plotting (limit to 6)
+        top_cats = merged['crime_category'].value_counts().nlargest(6).index.tolist()
+
+        # Plot distributions (KDE) of true and predicted per category
+        for cat in top_cats:
+            subset = merged[merged['crime_category'] == cat]
+            if len(subset) < 5:
+                print(f"Skipping category '{cat}' (only {len(subset)} samples)")
+                continue
+
+            plt.figure(figsize=(8, 4))
+            try:
+                sns.kdeplot(subset['true_label'], label='true', fill=True)
+            except Exception:
+                plt.hist(subset['true_label'], bins=30, alpha=0.4, density=True, label='true')
+            try:
+                sns.kdeplot(subset['prediction'], label='pred', color='orange', fill=True)
+            except Exception:
+                plt.hist(subset['prediction'], bins=30, alpha=0.4, density=True, label='pred', color='orange')
+
+            plt.title(f'Distribution for category: {cat}')
+            plt.xlabel('Days')
+            plt.legend()
+            plt.tight_layout()
+            fcat = img_dir / f"dist_{re.sub(r'[^a-zA-Z0-9]+','_', cat)[:50]}.png"
+            plt.savefig(fcat)
+            plt.close()
+            print(f"  - Saved distribution for '{cat}': {fcat}")
+
+        # Additionally: combined faceted histograms for top categories
+        try:
+            combined = merged[merged['crime_category'].isin(top_cats)].copy()
+            combined = combined.melt(id_vars=['file_name', 'crime_category'], value_vars=['true_label', 'prediction'], var_name='which', value_name='days')
+            g = sns.FacetGrid(combined, col='crime_category', hue='which', sharex=False, sharey=False, col_wrap=3)
+            g.map(sns.histplot, 'days', bins=30, alpha=0.6)
+            g.add_legend()
+            combined_plot = img_dir / 'combined_category_histograms.png'
+            plt.tight_layout()
+            plt.savefig(combined_plot)
+            plt.close()
+            print(f"✅ Combined category histograms saved: {combined_plot}")
+        except Exception as e:
+            print("⚠️ Failed to create combined faceted histograms:", e)
+    except Exception as e:
+        print("⚠️ Could not load test metadata for per-category plots:", e)
+except Exception as e:
+    print("⚠️ Plotting failed:", e)
 
 
 # ======================================================

32 Indobert-base-p1_with_law_with_svr.py

@@ -7,6 +7,7 @@
 import os
 import re
 import json
+import sys
 import zipfile
 import shutil
 import tempfile
@@ -258,9 +259,18 @@ tokenizer.save_pretrained(OUTPUT_DIR)
 
 
 # ======================================================
-# 6. Evaluation
+# 6. Evaluation (Comprehensive)
 # ======================================================
-print("\n📊 Evaluating on Test Set...")
+from sklearn.metrics import (
+    mean_squared_error, 
+    mean_absolute_error, 
+    r2_score,
+    mean_absolute_percentage_error
+)
+
+print("\n" + "="*70)
+print("📊 COMPUTING COMPREHENSIVE TEST SET METRICS")
+print("="*70)
 
 # Predict (Output is scaled)
 preds_scaled = svr_pipeline.predict(X_test)
@@ -272,50 +282,268 @@ y_true_days = label_scaler.inverse_transform(y_test_scaled.reshape(-1, 1)).flatt
 # Clip negatives (sentences can't be negative)
 preds_days = np.maximum(preds_days, 0)
 
-# Metrics
+# Basic metrics
 mse = mean_squared_error(y_true_days, preds_days)
 rmse = np.sqrt(mse)
 mae = mean_absolute_error(y_true_days, preds_days)
 r2 = r2_score(y_true_days, preds_days)
 
-print("\n🎯 Test Results:")
-print(f"  RMSE: {rmse:,.2f} days")
-print(f"  MAE:  {mae:,.2f} days")
-print(f"  R2:   {r2:.4f}")
+# Additional metrics
+median_ae = np.median(np.abs(y_true_days - preds_days))
+max_error = np.max(np.abs(y_true_days - preds_days))
+min_error = np.min(np.abs(y_true_days - preds_days))
+
+# Percentage errors (avoid division by zero)
+mask = y_true_days != 0
+if mask.sum() > 0:
+    mape = np.mean(np.abs((y_true_days[mask] - preds_days[mask]) / y_true_days[mask])) * 100
+else:
+    mape = float('nan')
+
+# Statistics
+pred_mean = np.mean(preds_days)
+pred_std = np.std(preds_days)
+pred_min = np.min(preds_days)
+pred_max = np.max(preds_days)
+
+true_mean = np.mean(y_true_days)
+true_std = np.std(y_true_days)
+true_min = np.min(y_true_days)
+true_max = np.max(y_true_days)
+
+print("\n🎯 ERROR METRICS:")
+print(f"  MSE:                {mse:>12,.2f} days²")
+print(f"  RMSE:               {rmse:>12,.2f} days")
+print(f"  MAE:                {mae:>12,.2f} days")
+print(f"  Median AE:          {median_ae:>12,.2f} days")
+print(f"  Max Error:          {max_error:>12,.2f} days")
+print(f"  Min Error:          {min_error:>12,.2f} days")
+if not np.isnan(mape):
+    print(f"  MAPE:               {mape:>12,.2f}%")
+print(f"  R² Score:           {r2:>12,.4f}")
+
+print("\n📊 PREDICTION STATISTICS:")
+print(f"  Mean:               {pred_mean:>12,.2f} days")
+print(f"  Std Dev:            {pred_std:>12,.2f} days")
+print(f"  Min:                {pred_min:>12,.2f} days")
+print(f"  Max:                {pred_max:>12,.2f} days")
+
+print("\n📊 TRUE LABEL STATISTICS:")
+print(f"  Mean:               {true_mean:>12,.2f} days")
+print(f"  Std Dev:            {true_std:>12,.2f} days")
+print(f"  Min:                {true_min:>12,.2f} days")
+print(f"  Max:                {true_max:>12,.2f} days")
+
+# Error distribution analysis
+print("\n📉 ERROR DISTRIBUTION:")
+errors = np.abs(y_true_days - preds_days)
+percentiles = [10, 25, 50, 75, 90, 95, 99]
+print("  Percentiles of Absolute Error:")
+for p in percentiles:
+    val = np.percentile(errors, p)
+    print(f"    {p}th percentile:   {val:>12,.2f} days")
+
+# Accuracy within ranges
+print("\n🎯 ACCURACY WITHIN ERROR RANGES:")
+for threshold in [100, 250, 500, 750, 1000]:
+    within = np.sum(errors <= threshold) / len(errors) * 100
+    print(f"  Within ±{threshold:>4} days:  {within:>6.2f}% of samples")
 
-# Save detailed results
+
+# ======================================================
+# SAVE DETAILED RESULTS
+# ======================================================
+detailed_results = {
+    'error_metrics': {
+        'mse': float(mse),
+        'rmse': float(rmse),
+        'mae': float(mae),
+        'median_ae': float(median_ae),
+        'max_error': float(max_error),
+        'min_error': float(min_error),
+        'mape': float(mape) if not np.isnan(mape) else None,
+        'r2_score': float(r2)
+    },
+    'prediction_stats': {
+        'mean': float(pred_mean),
+        'std': float(pred_std),
+        'min': float(pred_min),
+        'max': float(pred_max)
+    },
+    'true_label_stats': {
+        'mean': float(true_mean),
+        'std': float(true_std),
+        'min': float(true_min),
+        'max': float(true_max)
+    },
+    'error_percentiles': {
+        f'p{p}': float(np.percentile(errors, p)) for p in percentiles
+    },
+    'accuracy_within_ranges': {
+        f'within_{threshold}': float(np.sum(errors <= threshold) / len(errors) * 100)
+        for threshold in [100, 250, 500, 750, 1000]
+    }
+}
+
+results_detailed_file = Path(OUTPUT_DIR) / 'test_results_detailed.json'
+with open(results_detailed_file, 'w') as f:
+    json.dump(detailed_results, f, indent=2)
+
+print(f"\n💾 Detailed results saved to: {results_detailed_file}")
+
+# Save predictions vs true labels CSV
 results_df = pd.DataFrame({
     'file_name': test_filenames,
     'true_label': y_true_days,
     'prediction': preds_days,
-    'error': np.abs(y_true_days - preds_days)
+    'absolute_error': errors,
+    'percentage_error': np.abs((y_true_days - preds_days) / (y_true_days + 1)) * 100  # +1 to avoid div by 0
 })
-results_csv = Path(OUTPUT_DIR) / 'svr_test_predictions.csv'
+results_csv = Path(OUTPUT_DIR) / 'test_predictions.csv'
 results_df.to_csv(results_csv, index=False)
-print(f"💾 Detailed predictions saved to {results_csv}")
+print(f"💾 Predictions CSV saved to: {results_csv}")
+
+# Also save a simple inference CSV with filename and predicted value (user-friendly export)
+try:
+    # Ensure alignment in case of any mismatch
+    n_simple = min(len(test_filenames), len(preds_days))
+    simple_df = pd.DataFrame({
+        'filename': test_filenames[:n_simple],
+        'predicted_day_sentece': np.array(preds_days[:n_simple]).flatten()
+    })
+    simple_results_file = Path(OUTPUT_DIR) / 'test_inference_simple.csv'
+    simple_df.to_csv(simple_results_file, index=False)
+    print(f"💾 Simple inference CSV saved to: {simple_results_file}")
+except Exception as e:
+    print(f"⚠️ Failed to save simple inference CSV: {e}")
+
+# ======================================================
+# 🧪 SAMPLE PREDICTIONS DISPLAY
+# ======================================================
+print("\n" + "="*70)
+print("🧪 SAMPLE PREDICTIONS (First 10)")
+print("="*70)
+for i in range(min(10, len(preds_days))):
+    fname = test_filenames[i] if i < len(test_filenames) else ''
+    true_val = y_true_days[i]
+    pred_val = preds_days[i]
+    error = abs(true_val - pred_val)
+    pct_error = (error / (true_val + 1)) * 100
+    
+    print(f"\nSample {i+1}:")
+    print(f"  File:       {fname}")
+    print(f"  True:       {true_val:>8,.0f} days")
+    print(f"  Predicted:  {pred_val:>8,.0f} days")
+    print(f"  Error:      {error:>8,.0f} days ({pct_error:.1f}%)")
+
+print("\n" + "="*70)
+print("✅ COMPREHENSIVE TEST SET EVALUATION COMPLETE!")
+print(f"   - {results_detailed_file}")
+print(f"   - {results_csv}")
+print(f"   - {OUTPUT_DIR}/label_scaler.pkl")
+print(f"   - {OUTPUT_DIR}/svr_model.pkl")
 
 
 # ======================================================
 # 7. Plotting
 # ======================================================
-print("\n📈 Generating Plots...")
-img_dir = Path(OUTPUT_DIR) / 'plots'
-img_dir.mkdir(parents=True, exist_ok=True)
-
-# Scatter Plot
-plt.figure(figsize=(8, 8))
-plt.scatter(results_df['true_label'], results_df['prediction'], alpha=0.5, color='purple')
-lims = [0, max(results_df['true_label'].max(), results_df['prediction'].max()) * 1.05]
-plt.plot(lims, lims, '--', color='gray')
-plt.xlim(lims)
-plt.ylim(lims)
-plt.title(f'SVR: True vs Predicted Days (R2={r2:.3f})')
-plt.xlabel('True Days')
-plt.ylabel('Predicted Days')
-plt.grid(True, alpha=0.3)
-plt.tight_layout()
-plt.savefig(img_dir / 'svr_scatter.png')
-print(f"🖼️ Plot saved to {img_dir / 'svr_scatter.png'}")
+try:
+    print("\n🔍 Generating plots for predictions...")
+    img_dir = Path(OUTPUT_DIR) / 'plots'
+    img_dir.mkdir(parents=True, exist_ok=True)
+
+    # Basic scatterplots over sample index
+    idx = np.arange(len(results_df))
+
+    plt.figure(figsize=(12, 4))
+    plt.scatter(idx, results_df['true_label'], s=6, alpha=0.6)
+    plt.title('Ground truth (true_label) over samples')
+    plt.xlabel('Sample index')
+    plt.ylabel('Days')
+    plt.tight_layout()
+    p1 = img_dir / 'scatter_true.png'
+    plt.savefig(p1)
+    plt.close()
+
+    plt.figure(figsize=(12, 4))
+    plt.scatter(idx, results_df['prediction'], s=6, alpha=0.6, color='orange')
+    plt.title('Predictions over samples')
+    plt.xlabel('Sample index')
+    plt.ylabel('Predicted days')
+    plt.tight_layout()
+    p2 = img_dir / 'scatter_pred.png'
+    plt.savefig(p2)
+    plt.close()
+
+    # True vs Predicted scatter (combination)
+    plt.figure(figsize=(6, 6))
+    plt.scatter(results_df['true_label'], results_df['prediction'], s=8, alpha=0.5)
+    lims = [0, max(results_df['true_label'].max(), results_df['prediction'].max()) * 1.05]
+    plt.plot(lims, lims, '--', color='gray')
+    plt.xlim(lims)
+    plt.ylim(lims)
+    plt.xlabel('True days')
+    plt.ylabel('Predicted days')
+    plt.title('True vs Predicted')
+    plt.tight_layout()
+    p3 = img_dir / 'scatter_true_vs_pred.png'
+    plt.savefig(p3)
+    plt.close()
+
+    print(f"✅ Scatter plots saved: {p1}, {p2}, {p3}")
+
+    # Merge with test metadata to get crime_category (charge classification)
+    try:
+        meta_test = pd.read_csv(splits_dir / 'test_split.csv')
+        merged = results_df.merge(meta_test[['file_name', 'crime_category']], on='file_name', how='left')
+
+        # Choose top categories by sample count for plotting (limit to 6)
+        top_cats = merged['crime_category'].value_counts().nlargest(6).index.tolist()
+
+        # Plot distributions (KDE) of true and predicted per category
+        for cat in top_cats:
+            subset = merged[merged['crime_category'] == cat]
+            if len(subset) < 5:
+                print(f"Skipping category '{cat}' (only {len(subset)} samples)")
+                continue
+
+            plt.figure(figsize=(8, 4))
+            try:
+                sns.kdeplot(subset['true_label'], label='true', fill=True)
+            except Exception:
+                plt.hist(subset['true_label'], bins=30, alpha=0.4, density=True, label='true')
+            try:
+                sns.kdeplot(subset['prediction'], label='pred', color='orange', fill=True)
+            except Exception:
+                plt.hist(subset['prediction'], bins=30, alpha=0.4, density=True, label='pred', color='orange')
+
+            plt.title(f'Distribution for category: {cat}')
+            plt.xlabel('Days')
+            plt.legend()
+            plt.tight_layout()
+            fcat = img_dir / f"dist_{re.sub(r'[^a-zA-Z0-9]+','_', cat)[:50]}.png"
+            plt.savefig(fcat)
+            plt.close()
+            print(f"  - Saved distribution for '{cat}': {fcat}")
+
+        # Additionally: combined faceted histograms for top categories
+        try:
+            combined = merged[merged['crime_category'].isin(top_cats)].copy()
+            combined = combined.melt(id_vars=['file_name', 'crime_category'], value_vars=['true_label', 'prediction'], var_name='which', value_name='days')
+            g = sns.FacetGrid(combined, col='crime_category', hue='which', sharex=False, sharey=False, col_wrap=3)
+            g.map(sns.histplot, 'days', bins=30, alpha=0.6)
+            g.add_legend()
+            combined_plot = img_dir / 'combined_category_histograms.png'
+            plt.tight_layout()
+            plt.savefig(combined_plot)
+            plt.close()
+            print(f"✅ Combined category histograms saved: {combined_plot}")
+        except Exception as e:
+            print("⚠️ Failed to create combined faceted histograms:", e)
+    except Exception as e:
+        print("⚠️ Could not load test metadata for per-category plots:", e)
+except Exception as e:
+    print("⚠️ Plotting failed:", e)
 
 
 # ======================================================

51 Indobert-base-p1_without_law_with_xgboost.py

@@ -27,6 +27,29 @@ from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score
 
 # Import XGBoost
 import xgboost as xgb
+import logging
+from sklearn.metrics import mean_absolute_percentage_error
+
+# ======================================================
+# Logging Setup
+# ======================================================
+def setup_logging(output_dir):
+    Path(output_dir).mkdir(parents=True, exist_ok=True)
+    log_file = Path(output_dir) / 'training.log'
+    
+    # Remove existing handlers
+    for handler in logging.root.handlers[:]:
+        logging.root.removeHandler(handler)
+    
+    logging.basicConfig(
+        level=logging.INFO,
+        format='%(asctime)s - %(levelname)s - %(message)s',
+        handlers=[
+            logging.FileHandler(log_file, mode='w', encoding='utf-8'),
+            logging.StreamHandler()
+        ]
+    )
+    return logging.getLogger(__name__)
 
 # ======================================================
 # User Configuration
@@ -48,9 +71,15 @@ def set_seed(seed: int):
 
 set_seed(SEED)
 
+# Setup logging
+logger = setup_logging(OUTPUT_DIR)
+
 HF_TOKEN = os.environ.get('HF_TOKEN')  # Set HF_TOKEN environment variable
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-print(f"Device detected: {device}")
+logger.info(f"Device detected: {device}")
+logger.info(f"Model: {MODEL_NAME}")
+logger.info(f"Output directory: {OUTPUT_DIR}")
+logger.info(f"Max length: {MAX_LENGTH}, Batch size: {BATCH_SIZE}, Seed: {SEED}")
 
 # ======================================================
 # 1. Dataset Download & Prep
@@ -58,7 +87,7 @@ print(f"Device detected: {device}")
 REPO_ID = "evanslur/skripsi"
 ZIP_FILENAME = "dataset.zip"
 
-print("Downloading dataset...")
+logger.info("Downloading dataset...")
 zip_path = hf_hub_download(repo_id=REPO_ID, filename=ZIP_FILENAME, repo_type="dataset", token=HF_TOKEN)
 
 extract_dir = Path("./colloquial_data")
@@ -111,15 +140,17 @@ def preprocess_text(text):
     t = re.sub(r"\s+", ' ', t).strip().lower()
     return t
 
-print('Preprocessing text...')
+logger.info('Preprocessing text...')
 for split in loaded:
     loaded[split] = loaded[split].map(lambda x: {'text': preprocess_text(x['text'])}, num_proc=4)
 
+logger.info(f"Dataset loaded - Train: {len(loaded['train'])}, Val: {len(loaded['validation'])}, Test: {len(loaded['test'])}")
+
 # ======================================================
 # 3. Label Scaling
 # ======================================================
 # Meskipun Tree-Based model tidak wajib scaling, ini membantu normalisasi loss function
-print("Fitting StandardScaler...")
+logger.info("Fitting StandardScaler...")
 scaler = StandardScaler()
 train_labels_raw = np.array(loaded['train']['day_sentence']).reshape(-1, 1)
 scaler.fit(train_labels_raw)
@@ -128,14 +159,17 @@ Path(OUTPUT_DIR).mkdir(parents=True, exist_ok=True)
 with open(Path(OUTPUT_DIR) / 'scaler.pkl', 'wb') as f:
     pickle.dump(scaler, f)
 
+logger.info(f"Scaler fitted: mean={scaler.mean_[0]:.2f}, std={scaler.scale_[0]:.2f}")
+
 # ======================================================
 # 4. Feature Extraction (IndoBERT)
 # ======================================================
-print(f"Loading IndoBERT: {MODEL_NAME}")
+logger.info(f"Loading IndoBERT: {MODEL_NAME}")
 tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, token=HF_TOKEN)
 encoder = AutoModel.from_pretrained(MODEL_NAME, token=HF_TOKEN)
 encoder.to(device)
 encoder.eval()
+logger.info("IndoBERT encoder loaded successfully")
 
 def extract_embeddings(dataset_split, batch_size=32):
     all_embeddings = []
@@ -145,7 +179,7 @@ def extract_embeddings(dataset_split, batch_size=32):
     # Normalize labels
     labels_norm = scaler.transform(np.array(labels).reshape(-1, 1)).flatten()
     
-    print(f"Extracting features for {len(texts)} samples...")
+    logger.info(f"Extracting features for {len(texts)} samples...")
     for i in tqdm(range(0, len(texts), batch_size)):
         batch_texts = texts[i : i + batch_size]
         inputs = tokenizer(batch_texts, padding=True, truncation=True, max_length=MAX_LENGTH, return_tensors="pt").to(device)
@@ -164,11 +198,26 @@ X_test, y_test = extract_embeddings(loaded['test'], BATCH_SIZE)
 # ======================================================
 # 5. Train XGBoost (Tree Based Model)
 # ======================================================
-print("\n" + "="*50)
-print("🌳 Training XGBoost Regressor...")
-print("="*50)
+logger.info("\n" + "="*50)
+logger.info("🌳 Training XGBoost Regressor...")
+logger.info("="*50)
+logger.info(f"Feature dimension: {X_train.shape[1]}")
+logger.info(f"Training samples: {X_train.shape[0]}, Validation samples: {X_val.shape[0]}, Test samples: {X_test.shape[0]}")
 
 # Konfigurasi XGBoost
+xgb_params = {
+    'n_estimators': 1000,
+    'learning_rate': 0.05,
+    'max_depth': 6,
+    'subsample': 0.8,
+    'colsample_bytree': 0.8,
+    'objective': 'reg:squarederror',
+    'n_jobs': -1,
+    'random_state': SEED,
+    'early_stopping_rounds': 50
+}
+logger.info(f"XGBoost parameters: {xgb_params}")
+
 xgb_model = xgb.XGBRegressor(
     n_estimators=1000,      # Jumlah pohon
     learning_rate=0.05,     # Kecepatan belajar (makin kecil makin teliti tapi lambat)
@@ -188,10 +237,17 @@ xgb_model.fit(
     verbose=100
 )
 
-print("✅ XGBoost Training Completed!")
+logger.info("✅ XGBoost Training Completed!")
+logger.info(f"Best iteration: {xgb_model.best_iteration}")
+logger.info(f"Best score: {xgb_model.best_score:.6f}")
+
+# Get training history
+evals_result = xgb_model.evals_result()
+
 xgb_model.save_model(Path(OUTPUT_DIR) / "xgboost_model.json")
 encoder.save_pretrained(Path(OUTPUT_DIR) / "encoder")
 tokenizer.save_pretrained(OUTPUT_DIR)
+logger.info(f"Model saved to {OUTPUT_DIR}")
 
 # ======================================================
 # 6. Evaluation
@@ -207,13 +263,37 @@ def evaluate_model(model, X, y_true_norm, scaler, prefix="Test"):
     mae = mean_absolute_error(y_true_denorm, preds_denorm)
     r2 = r2_score(y_true_denorm, preds_denorm)
     
-    print(f"\n📊 {prefix} Results:")
-    print(f"  RMSE: {rmse:.2f} days")
-    print(f"  MAE:  {mae:.2f} days")
-    print(f"  R2:   {r2:.4f}")
-    return preds_denorm, y_true_denorm
+    # Additional metrics
+    median_ae = np.median(np.abs(y_true_denorm - preds_denorm))
+    max_error = np.max(np.abs(y_true_denorm - preds_denorm))
+    
+    # MAPE (avoid division by zero)
+    mask = y_true_denorm != 0
+    if mask.sum() > 0:
+        mape = mean_absolute_percentage_error(y_true_denorm[mask], preds_denorm[mask]) * 100
+    else:
+        mape = np.nan
+    
+    logger.info(f"\n📊 {prefix} Results:")
+    logger.info(f"  MSE:       {mse:.2f} days²")
+    logger.info(f"  RMSE:      {rmse:.2f} days")
+    logger.info(f"  MAE:       {mae:.2f} days")
+    logger.info(f"  Median AE: {median_ae:.2f} days")
+    logger.info(f"  Max Error: {max_error:.2f} days")
+    logger.info(f"  R2:        {r2:.4f}")
+    if not np.isnan(mape):
+        logger.info(f"  MAPE:      {mape:.2f}%")
+    
+    return preds_denorm, y_true_denorm, {'mse': mse, 'rmse': rmse, 'mae': mae, 'median_ae': median_ae, 'max_error': max_error, 'r2': r2, 'mape': mape}
 
-test_preds, test_true = evaluate_model(xgb_model, X_test, y_test, scaler, prefix="Test")
+# Evaluate all splits
+logger.info("\n" + "="*50)
+logger.info("📊 Evaluating on all splits...")
+logger.info("="*50)
+
+train_preds, train_true, train_metrics = evaluate_model(xgb_model, X_train, y_train, scaler, prefix="Train")
+val_preds, val_true, val_metrics = evaluate_model(xgb_model, X_val, y_val, scaler, prefix="Validation")
+test_preds, test_true, test_metrics = evaluate_model(xgb_model, X_test, y_test, scaler, prefix="Test")
 
 # ======================================================
 # 7. Save Results & Plot
@@ -227,24 +307,152 @@ results_df = pd.DataFrame({
 
 results_csv = Path(OUTPUT_DIR) / 'xgboost_test_predictions.csv'
 results_df.to_csv(results_csv, index=False)
-print(f"\n💾 Results saved to {results_csv}")
+logger.info(f"\n💾 Results saved to {results_csv}")
+
+# Save detailed results JSON
+detailed_results = {
+    'model': 'XGBoost',
+    'encoder': MODEL_NAME,
+    'dataset': 'colloquial (without law)',
+    'xgb_params': xgb_params,
+    'best_iteration': int(xgb_model.best_iteration),
+    'best_score': float(xgb_model.best_score),
+    'train_metrics': {k: float(v) if not np.isnan(v) else None for k, v in train_metrics.items()},
+    'val_metrics': {k: float(v) if not np.isnan(v) else None for k, v in val_metrics.items()},
+    'test_metrics': {k: float(v) if not np.isnan(v) else None for k, v in test_metrics.items()},
+    'prediction_stats': {
+        'mean': float(np.mean(test_preds)),
+        'std': float(np.std(test_preds)),
+        'min': float(np.min(test_preds)),
+        'max': float(np.max(test_preds))
+    },
+    'true_label_stats': {
+        'mean': float(np.mean(test_true)),
+        'std': float(np.std(test_true)),
+        'min': float(np.min(test_true)),
+        'max': float(np.max(test_true))
+    },
+    'error_percentiles': {
+        f'p{p}': float(np.percentile(results_df['error'], p)) for p in [10, 25, 50, 75, 90, 95, 99]
+    },
+    'accuracy_within_ranges': {
+        f'within_{threshold}': float(np.sum(results_df['error'] <= threshold) / len(results_df) * 100)
+        for threshold in [100, 250, 500, 750, 1000]
+    }
+}
+
+with open(Path(OUTPUT_DIR) / 'test_results_detailed.json', 'w') as f:
+    json.dump(detailed_results, f, indent=2)
+logger.info(f"Detailed results saved to {Path(OUTPUT_DIR) / 'test_results_detailed.json'}")
+
+# ======================================================
+# 8. Plots
+# ======================================================
+logger.info("\n" + "="*50)
+logger.info("🖼️ Generating plots...")
+logger.info("="*50)
 
-# Plot
 img_dir = Path(OUTPUT_DIR) / 'plots'
 img_dir.mkdir(parents=True, exist_ok=True)
-plt.figure(figsize=(6, 6))
-plt.scatter(results_df['true_label'], results_df['prediction'], alpha=0.5, color='green')
+
+# Plot 1: Scatter plot - True vs Predicted
+plt.figure(figsize=(8, 8))
+plt.scatter(results_df['true_label'], results_df['prediction'], alpha=0.5, color='green', edgecolors='none', s=30)
 lims = [0, max(results_df['true_label'].max(), results_df['prediction'].max()) * 1.05]
-plt.plot(lims, lims, '--', color='gray')
-plt.title(f'XGBoost: True vs Predicted')
-plt.xlabel('True Days')
-plt.ylabel('Predicted Days')
+plt.plot(lims, lims, '--', color='red', linewidth=2, label='Perfect Prediction')
+plt.title(f'XGBoost: True vs Predicted\nRMSE={test_metrics["rmse"]:.2f}, MAE={test_metrics["mae"]:.2f}, R²={test_metrics["r2"]:.4f}', fontsize=12)
+plt.xlabel('True Days', fontsize=11)
+plt.ylabel('Predicted Days', fontsize=11)
+plt.legend()
+plt.grid(True, alpha=0.3)
+plt.tight_layout()
+plt.savefig(img_dir / 'scatter_true_vs_predicted.png', dpi=150)
+plt.close()
+logger.info(f"  Saved: scatter_true_vs_predicted.png")
+
+# Plot 2: Residual Plot
+plt.figure(figsize=(10, 6))
+residuals = test_preds - test_true
+plt.scatter(test_true, residuals, alpha=0.5, color='blue', edgecolors='none', s=30)
+plt.axhline(y=0, color='red', linestyle='--', linewidth=2)
+plt.title('Residual Plot: Predicted - True', fontsize=12)
+plt.xlabel('True Days', fontsize=11)
+plt.ylabel('Residual (Predicted - True)', fontsize=11)
+plt.grid(True, alpha=0.3)
+plt.tight_layout()
+plt.savefig(img_dir / 'residual_plot.png', dpi=150)
+plt.close()
+logger.info(f"  Saved: residual_plot.png")
+
+# Plot 3: Error Distribution Histogram
+plt.figure(figsize=(10, 6))
+plt.hist(results_df['error'], bins=50, color='steelblue', edgecolor='white', alpha=0.8)
+plt.axvline(x=test_metrics['mae'], color='red', linestyle='--', linewidth=2, label=f'MAE = {test_metrics["mae"]:.2f}')
+plt.axvline(x=test_metrics['median_ae'], color='orange', linestyle='--', linewidth=2, label=f'Median AE = {test_metrics["median_ae"]:.2f}')
+plt.title('Distribution of Absolute Errors', fontsize=12)
+plt.xlabel('Absolute Error (Days)', fontsize=11)
+plt.ylabel('Frequency', fontsize=11)
+plt.legend()
+plt.grid(True, alpha=0.3)
+plt.tight_layout()
+plt.savefig(img_dir / 'error_distribution.png', dpi=150)
+plt.close()
+logger.info(f"  Saved: error_distribution.png")
+
+# Plot 4: Training Curve (XGBoost Loss)
+if evals_result:
+    plt.figure(figsize=(10, 6))
+    train_rmse = evals_result['validation_0']['rmse']
+    val_rmse = evals_result['validation_1']['rmse']
+    epochs = range(1, len(train_rmse) + 1)
+    plt.plot(epochs, train_rmse, 'b-', label='Train RMSE', linewidth=2)
+    plt.plot(epochs, val_rmse, 'r-', label='Validation RMSE', linewidth=2)
+    plt.axvline(x=xgb_model.best_iteration, color='green', linestyle='--', linewidth=1.5, label=f'Best iteration: {xgb_model.best_iteration}')
+    plt.title('XGBoost Training Curve', fontsize=12)
+    plt.xlabel('Boosting Round', fontsize=11)
+    plt.ylabel('RMSE (Normalized)', fontsize=11)
+    plt.legend()
+    plt.grid(True, alpha=0.3)
+    plt.tight_layout()
+    plt.savefig(img_dir / 'training_curve.png', dpi=150)
+    plt.close()
+    logger.info(f"  Saved: training_curve.png")
+
+# Plot 5: Box plot of errors by label range
+plt.figure(figsize=(12, 6))
+bins = [0, 365, 730, 1460, 2920, float('inf')]  # 0-1yr, 1-2yr, 2-4yr, 4-8yr, 8yr+
+labels_cat = ['0-1 year', '1-2 years', '2-4 years', '4-8 years', '8+ years']
+results_df['label_category'] = pd.cut(results_df['true_label'], bins=bins, labels=labels_cat)
+sns.boxplot(x='label_category', y='error', data=results_df, palette='viridis')
+plt.title('Error Distribution by Sentence Length Category', fontsize=12)
+plt.xlabel('True Sentence Category', fontsize=11)
+plt.ylabel('Absolute Error (Days)', fontsize=11)
+plt.xticks(rotation=15)
+plt.grid(True, alpha=0.3, axis='y')
 plt.tight_layout()
-plt.savefig(img_dir / 'xgboost_scatter.png')
-print(f"🖼️ Plot saved to {img_dir}")
+plt.savefig(img_dir / 'error_by_category_boxplot.png', dpi=150)
+plt.close()
+logger.info(f"  Saved: error_by_category_boxplot.png")
+
+# Plot 6: Feature Importance (Top 20)
+plt.figure(figsize=(10, 8))
+importances = xgb_model.feature_importances_
+top_k = 20
+top_indices = np.argsort(importances)[-top_k:][::-1]
+plt.barh(range(top_k), importances[top_indices], color='teal')
+plt.yticks(range(top_k), [f'Feature {i}' for i in top_indices])
+plt.xlabel('Importance', fontsize=11)
+plt.title(f'Top {top_k} Feature Importances (XGBoost)', fontsize=12)
+plt.gca().invert_yaxis()
+plt.tight_layout()
+plt.savefig(img_dir / 'feature_importance.png', dpi=150)
+plt.close()
+logger.info(f"  Saved: feature_importance.png")
+
+logger.info(f"🖼️ All plots saved to {img_dir}")
 
 # ======================================================
-# 8. Inference Function
+# 9. Inference Function
 # ======================================================
 def predict_sentence_tree(text):
     clean_text = preprocess_text(text)
@@ -258,4 +466,94 @@ def predict_sentence_tree(text):
     pred_days = scaler.inverse_transform(pred_norm.reshape(-1, 1))[0][0]
     return max(0, pred_days)
 
-print("\n🔮 Inference ready: predict_sentence_tree('text')")
+logger.info("\n🔮 Inference ready: predict_sentence_tree('text')")
+
+# ======================================================
+# 10. Final Summary
+# ======================================================
+logger.info("\n" + "="*70)
+logger.info("📈 TEST SET METRICS (Original Scale - Days)")
+logger.info("="*70)
+
+# Extract metrics
+mse = test_metrics['mse']
+rmse = test_metrics['rmse']
+mae = test_metrics['mae']
+median_ae = test_metrics['median_ae']
+max_error = test_metrics['max_error']
+r2 = test_metrics['r2']
+mape = test_metrics.get('mape', np.nan)
+min_error = np.min(results_df['error'])
+
+# Statistics
+pred_mean = np.mean(test_preds)
+pred_std = np.std(test_preds)
+pred_min = np.min(test_preds)
+pred_max = np.max(test_preds)
+
+true_mean = np.mean(test_true)
+true_std = np.std(test_true)
+true_min = np.min(test_true)
+true_max = np.max(test_true)
+
+logger.info("\n🎯 ERROR METRICS:")
+logger.info(f"  MSE:                {mse:>12,.2f} days²")
+logger.info(f"  RMSE:               {rmse:>12,.2f} days")
+logger.info(f"  MAE:                {mae:>12,.2f} days")
+logger.info(f"  Median AE:          {median_ae:>12,.2f} days")
+logger.info(f"  Max Error:          {max_error:>12,.2f} days")
+logger.info(f"  Min Error:          {min_error:>12,.2f} days")
+if not np.isnan(mape):
+    logger.info(f"  MAPE:               {mape:>12,.2f} %")
+logger.info(f"  R² Score:           {r2:>12,.4f}")
+
+logger.info("\n📊 PREDICTION STATISTICS:")
+logger.info(f"  Mean:               {pred_mean:>12,.2f} days")
+logger.info(f"  Std Dev:            {pred_std:>12,.2f} days")
+logger.info(f"  Min:                {pred_min:>12,.2f} days")
+logger.info(f"  Max:                {pred_max:>12,.2f} days")
+
+logger.info("\n📊 TRUE LABEL STATISTICS:")
+logger.info(f"  Mean:               {true_mean:>12,.2f} days")
+logger.info(f"  Std Dev:            {true_std:>12,.2f} days")
+logger.info(f"  Min:                {true_min:>12,.2f} days")
+logger.info(f"  Max:                {true_max:>12,.2f} days")
+
+# Error distribution analysis
+logger.info("\n📉 ERROR DISTRIBUTION:")
+percentiles = [10, 25, 50, 75, 90, 95, 99]
+logger.info("  Percentiles of Absolute Error:")
+for p in percentiles:
+    perc_val = np.percentile(results_df['error'], p)
+    logger.info(f"    {p}th percentile:   {perc_val:>12,.2f} days")
+
+# Accuracy within ranges
+logger.info("\n🎯 ACCURACY WITHIN ERROR RANGES:")
+for threshold in [100, 250, 500, 750, 1000]:
+    acc = np.sum(results_df['error'] <= threshold) / len(results_df) * 100
+    logger.info(f"    Within {threshold:>4} days: {acc:>12.2f}%")
+
+# Sample predictions
+logger.info("\n" + "="*70)
+logger.info("🧪 SAMPLE PREDICTIONS (First 10)")
+logger.info("="*70)
+for i in range(min(10, len(results_df))):
+    row = results_df.iloc[i]
+    logger.info(f"Sample {i+1}:")
+    logger.info(f"  File: {row['file_name']}")
+    logger.info(f"  True: {row['true_label']:.2f} days")
+    logger.info(f"  Pred: {row['prediction']:.2f} days")
+    logger.info(f"  Diff: {row['error']:.2f} days")
+    logger.info("-" * 30)
+
+logger.info("\n" + "="*70)
+logger.info("✅ TRAINING AND EVALUATION COMPLETE!")
+logger.info("="*70)
+logger.info(f"\nOutput files:")
+logger.info(f"  - Model: {Path(OUTPUT_DIR) / 'xgboost_model.json'}")
+logger.info(f"  - Encoder: {Path(OUTPUT_DIR) / 'encoder'}")
+logger.info(f"  - Scaler: {Path(OUTPUT_DIR) / 'scaler.pkl'}")
+logger.info(f"  - Predictions: {results_csv}")
+logger.info(f"  - Detailed results: {Path(OUTPUT_DIR) / 'test_results_detailed.json'}")
+logger.info(f"  - Training log: {Path(OUTPUT_DIR) / 'training.log'}")
+logger.info(f"  - Plots: {img_dir}")

52 Indobert-base-p1_with_law_with_xgboost.py

@@ -27,6 +27,29 @@ from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score
 
 # Import XGBoost
 import xgboost as xgb
+import logging
+from sklearn.metrics import mean_absolute_percentage_error
+
+# ======================================================
+# Logging Setup
+# ======================================================
+def setup_logging(output_dir):
+    Path(output_dir).mkdir(parents=True, exist_ok=True)
+    log_file = Path(output_dir) / 'training.log'
+    
+    # Remove existing handlers
+    for handler in logging.root.handlers[:]:
+        logging.root.removeHandler(handler)
+    
+    logging.basicConfig(
+        level=logging.INFO,
+        format='%(asctime)s - %(levelname)s - %(message)s',
+        handlers=[
+            logging.FileHandler(log_file, mode='w', encoding='utf-8'),
+            logging.StreamHandler()
+        ]
+    )
+    return logging.getLogger(__name__)
 
 # ======================================================
 # User Configuration
@@ -48,9 +71,15 @@ def set_seed(seed: int):
 
 set_seed(SEED)
 
+# Setup logging
+logger = setup_logging(OUTPUT_DIR)
+
 HF_TOKEN = os.environ.get('HF_TOKEN')  # Set HF_TOKEN environment variable
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-print(f"Device detected: {device}")
+logger.info(f"Device detected: {device}")
+logger.info(f"Model: {MODEL_NAME}")
+logger.info(f"Output directory: {OUTPUT_DIR}")
+logger.info(f"Max length: {MAX_LENGTH}, Batch size: {BATCH_SIZE}, Seed: {SEED}")
 
 # ======================================================
 # 1. Dataset Download & Prep
@@ -58,7 +87,7 @@ print(f"Device detected: {device}")
 REPO_ID = "evanslur/skripsi"
 ZIP_FILENAME = "dataset.zip"
 
-print("Downloading dataset...")
+logger.info("Downloading dataset...")
 zip_path = hf_hub_download(repo_id=REPO_ID, filename=ZIP_FILENAME, repo_type="dataset", token=HF_TOKEN)
 
 extract_dir = Path("./colloquial_data")
@@ -111,15 +140,17 @@ def preprocess_text(text):
     t = re.sub(r"\s+", ' ', t).strip().lower()
     return t
 
-print('Preprocessing text...')
+logger.info('Preprocessing text...')
 for split in loaded:
     loaded[split] = loaded[split].map(lambda x: {'text': preprocess_text(x['text'])}, num_proc=4)
 
+logger.info(f"Dataset loaded - Train: {len(loaded['train'])}, Val: {len(loaded['validation'])}, Test: {len(loaded['test'])}")
+
 # ======================================================
 # 3. Label Scaling
 # ======================================================
 # Meskipun Tree-Based model tidak wajib scaling, ini membantu normalisasi loss function
-print("Fitting StandardScaler...")
+logger.info("Fitting StandardScaler...")
 scaler = StandardScaler()
 train_labels_raw = np.array(loaded['train']['day_sentence']).reshape(-1, 1)
 scaler.fit(train_labels_raw)
@@ -128,14 +159,17 @@ Path(OUTPUT_DIR).mkdir(parents=True, exist_ok=True)
 with open(Path(OUTPUT_DIR) / 'scaler.pkl', 'wb') as f:
     pickle.dump(scaler, f)
 
+logger.info(f"Scaler fitted: mean={scaler.mean_[0]:.2f}, std={scaler.scale_[0]:.2f}")
+
 # ======================================================
 # 4. Feature Extraction (IndoBERT)
 # ======================================================
-print(f"Loading IndoBERT: {MODEL_NAME}")
+logger.info(f"Loading IndoBERT: {MODEL_NAME}")
 tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, token=HF_TOKEN)
 encoder = AutoModel.from_pretrained(MODEL_NAME, token=HF_TOKEN)
 encoder.to(device)
 encoder.eval()
+logger.info("IndoBERT encoder loaded successfully")
 
 def extract_embeddings(dataset_split, batch_size=32):
     all_embeddings = []
@@ -145,7 +179,7 @@ def extract_embeddings(dataset_split, batch_size=32):
     # Normalize labels
     labels_norm = scaler.transform(np.array(labels).reshape(-1, 1)).flatten()
     
-    print(f"Extracting features for {len(texts)} samples...")
+    logger.info(f"Extracting features for {len(texts)} samples...")
     for i in tqdm(range(0, len(texts), batch_size)):
         batch_texts = texts[i : i + batch_size]
         inputs = tokenizer(batch_texts, padding=True, truncation=True, max_length=MAX_LENGTH, return_tensors="pt").to(device)
@@ -164,11 +198,26 @@ X_test, y_test = extract_embeddings(loaded['test'], BATCH_SIZE)
 # ======================================================
 # 5. Train XGBoost (Tree Based Model)
 # ======================================================
-print("\n" + "="*50)
-print("🌳 Training XGBoost Regressor...")
-print("="*50)
+logger.info("\n" + "="*50)
+logger.info("🌳 Training XGBoost Regressor...")
+logger.info("="*50)
+logger.info(f"Feature dimension: {X_train.shape[1]}")
+logger.info(f"Training samples: {X_train.shape[0]}, Validation samples: {X_val.shape[0]}, Test samples: {X_test.shape[0]}")
 
 # Konfigurasi XGBoost
+xgb_params = {
+    'n_estimators': 1000,
+    'learning_rate': 0.05,
+    'max_depth': 6,
+    'subsample': 0.8,
+    'colsample_bytree': 0.8,
+    'objective': 'reg:squarederror',
+    'n_jobs': -1,
+    'random_state': SEED,
+    'early_stopping_rounds': 50
+}
+logger.info(f"XGBoost parameters: {xgb_params}")
+
 xgb_model = xgb.XGBRegressor(
     n_estimators=1000,      # Jumlah pohon
     learning_rate=0.05,     # Kecepatan belajar (makin kecil makin teliti tapi lambat)
@@ -188,10 +237,17 @@ xgb_model.fit(
     verbose=100
 )
 
-print("✅ XGBoost Training Completed!")
+logger.info("✅ XGBoost Training Completed!")
+logger.info(f"Best iteration: {xgb_model.best_iteration}")
+logger.info(f"Best score: {xgb_model.best_score:.6f}")
+
+# Get training history
+evals_result = xgb_model.evals_result()
+
 xgb_model.save_model(Path(OUTPUT_DIR) / "xgboost_model.json")
 encoder.save_pretrained(Path(OUTPUT_DIR) / "encoder")
 tokenizer.save_pretrained(OUTPUT_DIR)
+logger.info(f"Model saved to {OUTPUT_DIR}")
 
 # ======================================================
 # 6. Evaluation
@@ -207,13 +263,37 @@ def evaluate_model(model, X, y_true_norm, scaler, prefix="Test"):
     mae = mean_absolute_error(y_true_denorm, preds_denorm)
     r2 = r2_score(y_true_denorm, preds_denorm)
     
-    print(f"\n📊 {prefix} Results:")
-    print(f"  RMSE: {rmse:.2f} days")
-    print(f"  MAE:  {mae:.2f} days")
-    print(f"  R2:   {r2:.4f}")
-    return preds_denorm, y_true_denorm
+    # Additional metrics
+    median_ae = np.median(np.abs(y_true_denorm - preds_denorm))
+    max_error = np.max(np.abs(y_true_denorm - preds_denorm))
+    
+    # MAPE (avoid division by zero)
+    mask = y_true_denorm != 0
+    if mask.sum() > 0:
+        mape = mean_absolute_percentage_error(y_true_denorm[mask], preds_denorm[mask]) * 100
+    else:
+        mape = np.nan
+    
+    logger.info(f"\n📊 {prefix} Results:")
+    logger.info(f"  MSE:       {mse:.2f} days²")
+    logger.info(f"  RMSE:      {rmse:.2f} days")
+    logger.info(f"  MAE:       {mae:.2f} days")
+    logger.info(f"  Median AE: {median_ae:.2f} days")
+    logger.info(f"  Max Error: {max_error:.2f} days")
+    logger.info(f"  R2:        {r2:.4f}")
+    if not np.isnan(mape):
+        logger.info(f"  MAPE:      {mape:.2f}%")
+    
+    return preds_denorm, y_true_denorm, {'mse': mse, 'rmse': rmse, 'mae': mae, 'median_ae': median_ae, 'max_error': max_error, 'r2': r2, 'mape': mape}
 
-test_preds, test_true = evaluate_model(xgb_model, X_test, y_test, scaler, prefix="Test")
+# Evaluate all splits
+logger.info("\n" + "="*50)
+logger.info("📊 Evaluating on all splits...")
+logger.info("="*50)
+
+train_preds, train_true, train_metrics = evaluate_model(xgb_model, X_train, y_train, scaler, prefix="Train")
+val_preds, val_true, val_metrics = evaluate_model(xgb_model, X_val, y_val, scaler, prefix="Validation")
+test_preds, test_true, test_metrics = evaluate_model(xgb_model, X_test, y_test, scaler, prefix="Test")
 
 # ======================================================
 # 7. Save Results & Plot
@@ -227,24 +307,152 @@ results_df = pd.DataFrame({
 
 results_csv = Path(OUTPUT_DIR) / 'xgboost_test_predictions.csv'
 results_df.to_csv(results_csv, index=False)
-print(f"\n💾 Results saved to {results_csv}")
+logger.info(f"\n💾 Results saved to {results_csv}")
+
+# Save detailed results JSON
+detailed_results = {
+    'model': 'XGBoost',
+    'encoder': MODEL_NAME,
+    'dataset': 'colloquial_with_law',
+    'xgb_params': xgb_params,
+    'best_iteration': int(xgb_model.best_iteration),
+    'best_score': float(xgb_model.best_score),
+    'train_metrics': {k: float(v) if not np.isnan(v) else None for k, v in train_metrics.items()},
+    'val_metrics': {k: float(v) if not np.isnan(v) else None for k, v in val_metrics.items()},
+    'test_metrics': {k: float(v) if not np.isnan(v) else None for k, v in test_metrics.items()},
+    'prediction_stats': {
+        'mean': float(np.mean(test_preds)),
+        'std': float(np.std(test_preds)),
+        'min': float(np.min(test_preds)),
+        'max': float(np.max(test_preds))
+    },
+    'true_label_stats': {
+        'mean': float(np.mean(test_true)),
+        'std': float(np.std(test_true)),
+        'min': float(np.min(test_true)),
+        'max': float(np.max(test_true))
+    },
+    'error_percentiles': {
+        f'p{p}': float(np.percentile(results_df['error'], p)) for p in [10, 25, 50, 75, 90, 95, 99]
+    },
+    'accuracy_within_ranges': {
+        f'within_{threshold}': float(np.sum(results_df['error'] <= threshold) / len(results_df) * 100)
+        for threshold in [100, 250, 500, 750, 1000]
+    }
+}
+
+with open(Path(OUTPUT_DIR) / 'test_results_detailed.json', 'w') as f:
+    json.dump(detailed_results, f, indent=2)
+logger.info(f"Detailed results saved to {Path(OUTPUT_DIR) / 'test_results_detailed.json'}")
+
+# ======================================================
+# 8. Plots
+# ======================================================
+logger.info("\n" + "="*50)
+logger.info("🖼️ Generating plots...")
+logger.info("="*50)
 
-# Plot
 img_dir = Path(OUTPUT_DIR) / 'plots'
 img_dir.mkdir(parents=True, exist_ok=True)
-plt.figure(figsize=(6, 6))
-plt.scatter(results_df['true_label'], results_df['prediction'], alpha=0.5, color='green')
+
+# Plot 1: Scatter plot - True vs Predicted
+plt.figure(figsize=(8, 8))
+plt.scatter(results_df['true_label'], results_df['prediction'], alpha=0.5, color='green', edgecolors='none', s=30)
 lims = [0, max(results_df['true_label'].max(), results_df['prediction'].max()) * 1.05]
-plt.plot(lims, lims, '--', color='gray')
-plt.title(f'XGBoost: True vs Predicted')
-plt.xlabel('True Days')
-plt.ylabel('Predicted Days')
+plt.plot(lims, lims, '--', color='red', linewidth=2, label='Perfect Prediction')
+plt.title(f'XGBoost: True vs Predicted\nRMSE={test_metrics["rmse"]:.2f}, MAE={test_metrics["mae"]:.2f}, R²={test_metrics["r2"]:.4f}', fontsize=12)
+plt.xlabel('True Days', fontsize=11)
+plt.ylabel('Predicted Days', fontsize=11)
+plt.legend()
+plt.grid(True, alpha=0.3)
+plt.tight_layout()
+plt.savefig(img_dir / 'scatter_true_vs_predicted.png', dpi=150)
+plt.close()
+logger.info(f"  Saved: scatter_true_vs_predicted.png")
+
+# Plot 2: Residual Plot
+plt.figure(figsize=(10, 6))
+residuals = test_preds - test_true
+plt.scatter(test_true, residuals, alpha=0.5, color='blue', edgecolors='none', s=30)
+plt.axhline(y=0, color='red', linestyle='--', linewidth=2)
+plt.title('Residual Plot: Predicted - True', fontsize=12)
+plt.xlabel('True Days', fontsize=11)
+plt.ylabel('Residual (Predicted - True)', fontsize=11)
+plt.grid(True, alpha=0.3)
+plt.tight_layout()
+plt.savefig(img_dir / 'residual_plot.png', dpi=150)
+plt.close()
+logger.info(f"  Saved: residual_plot.png")
+
+# Plot 3: Error Distribution Histogram
+plt.figure(figsize=(10, 6))
+plt.hist(results_df['error'], bins=50, color='steelblue', edgecolor='white', alpha=0.8)
+plt.axvline(x=test_metrics['mae'], color='red', linestyle='--', linewidth=2, label=f'MAE = {test_metrics["mae"]:.2f}')
+plt.axvline(x=test_metrics['median_ae'], color='orange', linestyle='--', linewidth=2, label=f'Median AE = {test_metrics["median_ae"]:.2f}')
+plt.title('Distribution of Absolute Errors', fontsize=12)
+plt.xlabel('Absolute Error (Days)', fontsize=11)
+plt.ylabel('Frequency', fontsize=11)
+plt.legend()
+plt.grid(True, alpha=0.3)
+plt.tight_layout()
+plt.savefig(img_dir / 'error_distribution.png', dpi=150)
+plt.close()
+logger.info(f"  Saved: error_distribution.png")
+
+# Plot 4: Training Curve (XGBoost Loss)
+if evals_result:
+    plt.figure(figsize=(10, 6))
+    train_rmse = evals_result['validation_0']['rmse']
+    val_rmse = evals_result['validation_1']['rmse']
+    epochs = range(1, len(train_rmse) + 1)
+    plt.plot(epochs, train_rmse, 'b-', label='Train RMSE', linewidth=2)
+    plt.plot(epochs, val_rmse, 'r-', label='Validation RMSE', linewidth=2)
+    plt.axvline(x=xgb_model.best_iteration, color='green', linestyle='--', linewidth=1.5, label=f'Best iteration: {xgb_model.best_iteration}')
+    plt.title('XGBoost Training Curve', fontsize=12)
+    plt.xlabel('Boosting Round', fontsize=11)
+    plt.ylabel('RMSE (Normalized)', fontsize=11)
+    plt.legend()
+    plt.grid(True, alpha=0.3)
+    plt.tight_layout()
+    plt.savefig(img_dir / 'training_curve.png', dpi=150)
+    plt.close()
+    logger.info(f"  Saved: training_curve.png")
+
+# Plot 5: Box plot of errors by label range
+plt.figure(figsize=(12, 6))
+bins = [0, 365, 730, 1460, 2920, float('inf')]  # 0-1yr, 1-2yr, 2-4yr, 4-8yr, 8yr+
+labels_cat = ['0-1 year', '1-2 years', '2-4 years', '4-8 years', '8+ years']
+results_df['label_category'] = pd.cut(results_df['true_label'], bins=bins, labels=labels_cat)
+sns.boxplot(x='label_category', y='error', data=results_df, palette='viridis')
+plt.title('Error Distribution by Sentence Length Category', fontsize=12)
+plt.xlabel('True Sentence Category', fontsize=11)
+plt.ylabel('Absolute Error (Days)', fontsize=11)
+plt.xticks(rotation=15)
+plt.grid(True, alpha=0.3, axis='y')
 plt.tight_layout()
-plt.savefig(img_dir / 'xgboost_scatter.png')
-print(f"🖼️ Plot saved to {img_dir}")
+plt.savefig(img_dir / 'error_by_category_boxplot.png', dpi=150)
+plt.close()
+logger.info(f"  Saved: error_by_category_boxplot.png")
+
+# Plot 6: Feature Importance (Top 20)
+plt.figure(figsize=(10, 8))
+importances = xgb_model.feature_importances_
+top_k = 20
+top_indices = np.argsort(importances)[-top_k:][::-1]
+plt.barh(range(top_k), importances[top_indices], color='teal')
+plt.yticks(range(top_k), [f'Feature {i}' for i in top_indices])
+plt.xlabel('Importance', fontsize=11)
+plt.title(f'Top {top_k} Feature Importances (XGBoost)', fontsize=12)
+plt.gca().invert_yaxis()
+plt.tight_layout()
+plt.savefig(img_dir / 'feature_importance.png', dpi=150)
+plt.close()
+logger.info(f"  Saved: feature_importance.png")
+
+logger.info(f"🖼️ All plots saved to {img_dir}")
 
 # ======================================================
-# 8. Inference Function
+# 9. Inference Function
 # ======================================================
 def predict_sentence_tree(text):
     clean_text = preprocess_text(text)
@@ -258,4 +466,94 @@ def predict_sentence_tree(text):
     pred_days = scaler.inverse_transform(pred_norm.reshape(-1, 1))[0][0]
     return max(0, pred_days)
 
-print("\n🔮 Inference ready: predict_sentence_tree('text')")
+logger.info("\n🔮 Inference ready: predict_sentence_tree('text')")
+
+# ======================================================
+# 10. Final Summary
+# ======================================================
+logger.info("\n" + "="*70)
+logger.info("📈 TEST SET METRICS (Original Scale - Days)")
+logger.info("="*70)
+
+# Extract metrics
+mse = test_metrics['mse']
+rmse = test_metrics['rmse']
+mae = test_metrics['mae']
+median_ae = test_metrics['median_ae']
+max_error = test_metrics['max_error']
+r2 = test_metrics['r2']
+mape = test_metrics.get('mape', np.nan)
+min_error = np.min(results_df['error'])
+
+# Statistics
+pred_mean = np.mean(test_preds)
+pred_std = np.std(test_preds)
+pred_min = np.min(test_preds)
+pred_max = np.max(test_preds)
+
+true_mean = np.mean(test_true)
+true_std = np.std(test_true)
+true_min = np.min(test_true)
+true_max = np.max(test_true)
+
+logger.info("\n🎯 ERROR METRICS:")
+logger.info(f"  MSE:                {mse:>12,.2f} days²")
+logger.info(f"  RMSE:               {rmse:>12,.2f} days")
+logger.info(f"  MAE:                {mae:>12,.2f} days")
+logger.info(f"  Median AE:          {median_ae:>12,.2f} days")
+logger.info(f"  Max Error:          {max_error:>12,.2f} days")
+logger.info(f"  Min Error:          {min_error:>12,.2f} days")
+if not np.isnan(mape):
+    logger.info(f"  MAPE:               {mape:>12,.2f} %")
+logger.info(f"  R² Score:           {r2:>12,.4f}")
+
+logger.info("\n📊 PREDICTION STATISTICS:")
+logger.info(f"  Mean:               {pred_mean:>12,.2f} days")
+logger.info(f"  Std Dev:            {pred_std:>12,.2f} days")
+logger.info(f"  Min:                {pred_min:>12,.2f} days")
+logger.info(f"  Max:                {pred_max:>12,.2f} days")
+
+logger.info("\n📊 TRUE LABEL STATISTICS:")
+logger.info(f"  Mean:               {true_mean:>12,.2f} days")
+logger.info(f"  Std Dev:            {true_std:>12,.2f} days")
+logger.info(f"  Min:                {true_min:>12,.2f} days")
+logger.info(f"  Max:                {true_max:>12,.2f} days")
+
+# Error distribution analysis
+logger.info("\n📉 ERROR DISTRIBUTION:")
+percentiles = [10, 25, 50, 75, 90, 95, 99]
+logger.info("  Percentiles of Absolute Error:")
+for p in percentiles:
+    perc_val = np.percentile(results_df['error'], p)
+    logger.info(f"    {p}th percentile:   {perc_val:>12,.2f} days")
+
+# Accuracy within ranges
+logger.info("\n🎯 ACCURACY WITHIN ERROR RANGES:")
+for threshold in [100, 250, 500, 750, 1000]:
+    acc = np.sum(results_df['error'] <= threshold) / len(results_df) * 100
+    logger.info(f"    Within {threshold:>4} days: {acc:>12.2f}%")
+
+# Sample predictions
+logger.info("\n" + "="*70)
+logger.info("🧪 SAMPLE PREDICTIONS (First 10)")
+logger.info("="*70)
+for i in range(min(10, len(results_df))):
+    row = results_df.iloc[i]
+    logger.info(f"Sample {i+1}:")
+    logger.info(f"  File: {row['file_name']}")
+    logger.info(f"  True: {row['true_label']:.2f} days")
+    logger.info(f"  Pred: {row['prediction']:.2f} days")
+    logger.info(f"  Diff: {row['error']:.2f} days")
+    logger.info("-" * 30)
+
+logger.info("\n" + "="*70)
+logger.info("✅ TRAINING AND EVALUATION COMPLETE!")
+logger.info("="*70)
+logger.info(f"\nOutput files:")
+logger.info(f"  - Model: {Path(OUTPUT_DIR) / 'xgboost_model.json'}")
+logger.info(f"  - Encoder: {Path(OUTPUT_DIR) / 'encoder'}")
+logger.info(f"  - Scaler: {Path(OUTPUT_DIR) / 'scaler.pkl'}")
+logger.info(f"  - Predictions: {results_csv}")
+logger.info(f"  - Detailed results: {Path(OUTPUT_DIR) / 'test_results_detailed.json'}")
+logger.info(f"  - Training log: {Path(OUTPUT_DIR) / 'training.log'}")
+logger.info(f"  - Plots: {img_dir}")

requirements.txt

@@ -12,6 +12,7 @@ pyarrow>=21.0.0,<25.0.0
 matplotlib
 seaborn
 xgboost
+sys
 
 # NOTE: We intentionally do NOT pin `torch` here. Install `torch` separately
 # using the official PyTorch instructions for your CUDA / CPU setup:

run_experiments.py

@@ -5,6 +5,7 @@ Modify the 'scripts' list to include the files you want to run.
 
 import subprocess
 import sys
+import os
 import time
 from datetime import datetime
 
@@ -56,6 +57,14 @@ def main():
     
     for i, script in enumerate(scripts, 1):
         print(f"\n[{i}/{len(scripts)}]", end="")
+        
+        # Check if output directory exists
+        output_dir = script.replace('.py', '')
+        if os.path.exists(output_dir):
+            print(f" Skipping {script} (Output directory '{output_dir}' already exists)")
+            results.append((script, True, 0))
+            continue
+
         success, elapsed = run_script(script)
         results.append((script, success, elapsed))