diff --git a/notebooks/GoEmotions_DeBERTa_SIMPLIFIED_PARALLEL.ipynb b/notebooks/GoEmotions_DeBERTa_SIMPLIFIED_PARALLEL.ipynb
index 6ce8d8e..27990ad 100644
--- a/notebooks/GoEmotions_DeBERTa_SIMPLIFIED_PARALLEL.ipynb
+++ b/notebooks/GoEmotions_DeBERTa_SIMPLIFIED_PARALLEL.ipynb
@@ -1,4 +1,879 @@
-{
+# RESUME TRAINING WITH ALL IMPROVEMENTS
+
+# Now we can resume training with:
+# 1. Fixed AsymmetricLoss implementation
+# 2. Progress monitoring to detect stalls
+# 3. Automatic Google Drive backup every 15 minutes
+# 4. Disk quota management and auto-cleanup
+
+import subprocess, threading, os, time
+from datetime import datetime
+
+# Kill any existing processes first
+subprocess.run(['pkill', '-f', 'train_deberta_local'], capture_output=True)
+time.sleep(2)
+
+print("🚀 PHASE 1.5: IMPROVED PARALLEL TRAINING WITH BACKUP & MONITORING")
+print("=" * 70)
+print("IMPROVEMENTS:")
+print("✅ AsymmetricLoss fixed to maintain gradient flow")
+print("✅ Progress monitoring to auto-detect training stalls")
+print("✅ Auto-backup to Google Drive every 15 minutes")
+print("✅ Disk quota management with auto-cleanup")
+print("✅ NCCL optimizations for better multi-GPU communication")
+print("=" * 70)
+
+# RESUME COMBINED MODELS SINCE THESE WERE INTERRUPTED
+print("\n📍 RESUMING: Combined 0.7 (GPU0) + Combined 0.5 (GPU1) with BACKUP & MONITORING")
+t1 = threading.Thread(target=run_config_with_monitor_and_backup, args=(0, 'Combined_07_Parallel', False, 0.7))
+t2 = threading.Thread(target=run_config_with_monitor_and_backup, args=(1, 'Combined_05_Parallel', False, 0.5))
+t1.start()
+t2.start()
+t1.join()
+t2.join()
+
+# Run Combined 0.3 after the others are done
+print("\n📍 SINGLE: Combined 0.3 (GPU0) with BACKUP & MONITORING")
+run_config_with_monitor_and_backup(0, 'Combined_03_Parallel', False, 0.3)
+
+print("\n🎉 PHASE 1.5 PARALLEL COMPLETE with ALL IMPROVEMENTS!")
+print("📊 Outputs: ./outputs/parallel_Combined_07_Parallel/, ./outputs/parallel_Combined_05_Parallel/, etc.")
+print("💾 Backups saved to: drive:00_Projects/🎯 TechLabs-2025/Final_Project/TRAINING/GoEmotions-DeBERTa-Backup")
+print("📈 Run the analysis cell to compare F1@0.2 scores")
+print("⚠️ If training stalls, it will automatically recover or notify you")# IMPROVED RCLONE BACKUP FOR GOOGLE DRIVE
+
+# First, let's test the rclone connection to ensure proper setup
+test_rclone_script = """#!/bin/bash
+
+# Test rclone connection to Google Drive
+echo "🔍 Testing rclone connection to Google Drive..."
+
+# Check if rclone is installed
+if ! command -v rclone &> /dev/null; then
+    echo "❌ rclone not found. Please install it first."
+    exit 1
+fi
+
+# Check if drive remote exists
+if ! rclone listremotes | grep -q "drive:"; then
+    echo "❌ 'drive:' remote not found in rclone configuration."
+    echo "Please configure rclone with 'rclone config' command first."
+    exit 1
+fi
+
+# Try to list root directory
+echo "📁 Testing access to drive: remote..."
+if ! rclone lsd drive: &>/dev/null; then
+    echo "❌ Cannot access 'drive:' remote. Please check your rclone configuration."
+    exit 1
+fi
+
+# Try to access or create the backup directory
+TARGET_DIR="drive:00_Projects/🎯 TechLabs-2025/Final_Project/TRAINING/GoEmotions-DeBERTa-Backup"
+echo "📁 Testing access to target directory: $TARGET_DIR"
+
+if ! rclone lsd "$TARGET_DIR" &>/dev/null; then
+    echo "🔨 Target directory doesn't exist. Creating it now..."
+    if ! rclone mkdir "$TARGET_DIR" &>/dev/null; then
+        echo "❌ Failed to create target directory. Please check permissions and path."
+        exit 1
+    else
+        echo "✅ Target directory created successfully."
+    fi
+else
+    echo "✅ Target directory exists and is accessible."
+fi
+
+# Try writing a test file
+TEST_FILE="/tmp/rclone_test_$(date +%s).txt"
+echo "This is a test file. Created at $(date)" > "$TEST_FILE"
+
+echo "📤 Uploading test file to $TARGET_DIR..."
+if ! rclone copy "$TEST_FILE" "$TARGET_DIR" &>/dev/null; then
+    echo "❌ Failed to upload test file. Please check permissions."
+    rm -f "$TEST_FILE"
+    exit 1
+else
+    echo "✅ Test file uploaded successfully."
+    
+    # Try to read the file back
+    echo "📥 Verifying test file..."
+    if ! rclone ls "$TARGET_DIR/$(basename "$TEST_FILE")" &>/dev/null; then
+        echo "⚠️ Warning: File uploaded but not immediately visible. This may be normal for Google Drive."
+    else
+        echo "✅ File verified in target location."
+        # Clean up the test file
+        rclone delete "$TARGET_DIR/$(basename "$TEST_FILE")" &>/dev/null
+    fi
+fi
+
+rm -f "$TEST_FILE"
+echo "✅ rclone connection test completed successfully!"
+echo "✅ Backup system ready to use with target: $TARGET_DIR"
+"""
+
+# Create and run the test script
+with open("/home/user/goemotions-deberta/test_rclone.sh", "w") as f:
+    f.write(test_rclone_script)
+!chmod +x /home/user/goemotions-deberta/test_rclone.sh
+!bash /home/user/goemotions-deberta/test_rclone.sh
+
+# Now create the improved backup script
+backup_script = """#!/bin/bash
+
+# Backup script for GoEmotions-DeBERTa model outputs
+# Syncs training outputs to Google Drive to prevent disk quota issues
+
+# Target Google Drive folder - EXACT path from rclone config
+DRIVE_TARGET="drive:00_Projects/🎯 TechLabs-2025/Final_Project/TRAINING/GoEmotions-DeBERTa-Backup"
+
+# Local paths to backup
+MODEL_OUTPUTS="./outputs"
+MODEL_CACHE="./models"
+DATASET_CACHE="./data"
+LOGS="./logs"
+
+# Timestamp
+TIMESTAMP=$(date +"%Y-%m-%d_%H-%M-%S")
+echo "🔄 Starting backup at $TIMESTAMP"
+
+# Check if rclone is configured
+if ! rclone version &>/dev/null; then
+    echo "❌ rclone not found or not configured"
+    exit 1
+fi
+
+# Verify the remote exists and we can access it
+if ! rclone lsd "drive:" &>/dev/null; then
+    echo "❌ Cannot access 'drive:' remote. Check rclone configuration."
+    exit 1
+fi
+
+# Check/create directory structure
+echo "🔍 Verifying backup directory structure..."
+if ! rclone lsd "$DRIVE_TARGET" &>/dev/null; then
+    echo "🔨 Creating main backup directory..."
+    rclone mkdir "$DRIVE_TARGET"
+fi
+
+# Create subdirectories
+for dir in "outputs" "models" "data" "logs"; do
+    if ! rclone lsd "$DRIVE_TARGET/$dir" &>/dev/null; then
+        echo "🔨 Creating $DRIVE_TARGET/$dir"
+        rclone mkdir "$DRIVE_TARGET/$dir"
+    fi
+done
+
+# Backup eval reports first (highest value, smallest size)
+echo "📊 Backing up evaluation reports..."
+find "$MODEL_OUTPUTS" -name "eval_report.json" -type f | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    target_dir=$(dirname "$target_path")
+    
+    # Ensure parent directory exists
+    rclone mkdir "$target_dir" 2>/dev/null
+    
+    echo "   $file → $target_path"
+    rclone copy "$file" "$target_dir"
+done
+
+# Backup model weights (most important for resuming)
+echo "🤖 Backing up model weights..."
+find "$MODEL_OUTPUTS" -name "pytorch_model.bin" -o -name "model.safetensors" | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    target_dir=$(dirname "$target_path")
+    
+    # Ensure parent directory exists
+    rclone mkdir "$target_dir" 2>/dev/null
+    
+    echo "   $file → $target_path"
+    rclone copy "$file" "$target_dir"
+done
+
+# Backup model configs
+echo "⚙️ Backing up model configs..."
+find "$MODEL_OUTPUTS" -name "config.json" -o -name "special_tokens_map.json" -o -name "tokenizer_config.json" -type f | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    target_dir=$(dirname "$target_path")
+    
+    # Ensure parent directory exists
+    rclone mkdir "$target_dir" 2>/dev/null
+    
+    echo "   $file → $target_path"
+    rclone copy "$file" "$target_dir"
+done
+
+# Backup base model cache (smaller files first)
+echo "💾 Backing up model metadata..."
+find "$MODEL_CACHE" -name "*.json" -type f | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    target_dir=$(dirname "$target_path")
+    
+    rclone mkdir "$target_dir" 2>/dev/null
+    rclone copy "$file" "$target_dir"
+done
+
+# Backup tokenizer files
+echo "🔤 Backing up tokenizer files..."
+find "$MODEL_CACHE" -name "*.model" -o -name "*.txt" -type f | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    target_dir=$(dirname "$target_path")
+    
+    rclone mkdir "$target_dir" 2>/dev/null
+    rclone copy "$file" "$target_dir"
+done
+
+# Backup dataset files (important but can be large)
+echo "📊 Backing up dataset metadata..."
+find "$DATASET_CACHE" -name "*.json" -type f | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    target_dir=$(dirname "$target_path")
+    
+    rclone mkdir "$target_dir" 2>/dev/null
+    rclone copy "$file" "$target_dir"
+done
+
+# Backup logs
+echo "📝 Backing up logs..."
+rclone copy "$LOGS" "$DRIVE_TARGET/logs" --update
+
+# Record backup in log file
+echo "$TIMESTAMP: Backup completed successfully" >> "./backup_history.log"
+rclone copy "./backup_history.log" "$DRIVE_TARGET/"
+
+# Check disk space after backup
+FREE_SPACE=$(df -h . | awk 'NR==2 {print $4}')
+USED_PERCENT=$(df -h . | awk 'NR==2 {print $5}')
+echo "💾 Disk space after backup: $FREE_SPACE free ($USED_PERCENT used)"
+
+echo "✅ Backup completed at $(date)"
+"""
+
+# Auto backup script that runs during training
+auto_backup_script = """#!/bin/bash
+
+# Auto backup script that runs during training
+# Set to backup every 15 minutes to prevent disk quota issues
+
+BACKUP_INTERVAL=900  # 15 minutes
+BACKUP_SCRIPT="/home/user/goemotions-deberta/backup_to_drive.sh"
+
+echo "🔄 Starting automatic backup service at $(date)"
+echo "📁 Target: drive:00_Projects/🎯 TechLabs-2025/Final_Project/TRAINING/GoEmotions-DeBERTa-Backup"
+echo "⏱️ Backup interval: $BACKUP_INTERVAL seconds (15 minutes)"
+echo "🧹 Auto-cleanup: Enabled for disk usage >85%"
+
+while true; do
+    # Run backup
+    echo ""
+    echo "🔄 Running scheduled backup ($(date))"
+    bash "$BACKUP_SCRIPT"
+    
+    # Check for disk quota issues
+    FREE_SPACE=$(df -h . | awk 'NR==2 {print $4}')
+    USED_PERCENT=$(df -h . | awk 'NR==2 {print $5}')
+    USED_NUM=${USED_PERCENT%\%}
+    
+    if [ "$USED_NUM" -gt 85 ]; then
+        echo "⚠️ WARNING: High disk usage ($USED_PERCENT)"
+        echo "🧹 Cleaning old checkpoints after backup..."
+        # Find and list what will be removed first
+        echo "Finding old checkpoints to remove..."
+        find ./outputs -name "checkpoint-*" -type d | sort | head -n -5
+        # Then remove them
+        find ./outputs -name "checkpoint-*" -type d | sort | head -n -5 | xargs rm -rf
+        
+        # Check space after cleanup
+        FREE_SPACE=$(df -h . | awk 'NR==2 {print $4}')
+        USED_PERCENT=$(df -h . | awk 'NR==2 {print $5}')
+        echo "💾 Disk space after cleanup: $FREE_SPACE free ($USED_PERCENT used)"
+    fi
+    
+    echo "💤 Next backup in $(($BACKUP_INTERVAL / 60)) minutes ($(date -d "+$BACKUP_INTERVAL seconds"))"
+    sleep $BACKUP_INTERVAL
+done
+"""
+
+# Create the backup scripts
+with open("/home/user/goemotions-deberta/backup_to_drive.sh", "w") as f:
+    f.write(backup_script)
+!chmod +x /home/user/goemotions-deberta/backup_to_drive.sh
+
+with open("/home/user/goemotions-deberta/auto_backup.sh", "w") as f:
+    f.write(auto_backup_script)
+!chmod +x /home/user/goemotions-deberta/auto_backup.sh
+
+# Run an initial backup to verify everything works
+print("🚀 Running initial backup to verify Google Drive connection...")
+!bash /home/user/goemotions-deberta/backup_to_drive.sh
+
+import subprocess
+# Function to start auto backup
+def start_auto_backup():
+    """Start automatic backup to Google Drive"""
+    print("🔄 Starting automatic Google Drive backup...")
+    subprocess.Popen(
+        ["bash", "/home/user/goemotions-deberta/auto_backup.sh"], 
+        stdout=subprocess.PIPE, 
+        stderr=subprocess.STDOUT,
+        text=True,
+        bufsize=1
+    )
+    print("✅ Auto backup process started - will backup every 15 minutes")
+    print("📁 Backup target: drive:00_Projects/🎯 TechLabs-2025/Final_Project/TRAINING/GoEmotions-DeBERTa-Backup")
+
+print("\n✅ Google Drive backup system configured!")
+print("📋 Features:")
+print("1. Verified rclone connection to Google Drive")
+print("2. Backs up model checkpoints, configs, and evaluation results every 15 minutes")
+print("3. Ensures parent directories exist on Google Drive")
+print("4. Automatically cleans up old checkpoints after backup when disk usage is high")
+print("5. All outputs are saved to: drive:00_Projects/🎯 TechLabs-2025/Final_Project/TRAINING/GoEmotions-DeBERTa-Backup")
+
+# Function to run training with auto backup
+def run_config_with_monitor_and_backup(gpu_id, config_name, use_asym=False, ratio=None):
+    """Run training with progress monitoring and auto backup"""
+    from datetime import datetime
+    print(f"🚀 Starting {config_name} on GPU {gpu_id} at {datetime.now()}")
+    
+    # Start auto backup process
+    start_auto_backup()
+    
+    cmd = ['./run_with_monitor.sh', str(gpu_id), config_name]
+    
+    # Add standard args
+    cmd.extend([
+        '--model_type', 'deberta-v3-large',
+        '--per_device_train_batch_size', '4',
+        '--per_device_eval_batch_size', '8',
+        '--gradient_accumulation_steps', '4',
+        '--num_train_epochs', '2',
+        '--learning_rate', '3e-5',
+        '--lr_scheduler_type', 'cosine',
+        '--warmup_ratio', '0.15',
+        '--weight_decay', '0.01',
+        '--fp16',
+        '--max_length', '256',
+        '--max_train_samples', '20000',
+        '--max_eval_samples', '3000'
+    ])
+    
+    if use_asym:
+        cmd.append('--use_asymmetric_loss')
+    
+    if ratio is not None:
+        cmd.extend(['--use_combined_loss', '--loss_combination_ratio', str(ratio)])
+    
+    print(f"Command for {config_name}: {' '.join(cmd)}")
+    
+    # Run in subprocess with full output capture
+    process = subprocess.Popen(
+        cmd,
+        cwd='/home/user/goemotions-deberta',
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+        text=True,
+        bufsize=1
+    )
+    
+    # Read output in real-time
+    for line in iter(process.stdout.readline, ''):
+        timestamp = datetime.now().strftime('%H:%M:%S')
+        print(f"[{timestamp}] GPU {gpu_id} [{config_name}]: {line.strip()}")
+    
+    # Wait for completion and get return code
+    return_code = process.wait()
+    
+    # Run final backup after training completes
+    print(f"🔄 Running final backup for {config_name}...")
+    subprocess.run(['bash', '/home/user/goemotions-deberta/backup_to_drive.sh'], 
+                   check=False, stdout=subprocess.PIPE)
+    
+    print(f"✅ {config_name} complete on GPU {gpu_id} (return code: {return_code})")
+    return return_code# RESUME TRAINING WITH ALL IMPROVEMENTS + GOOGLE DRIVE BACKUP
+
+# Now we resume training with:
+# 1. Fixed AsymmetricLoss implementation
+# 2. Progress monitoring to detect stalls
+# 3. Disk quota management and auto-cleanup
+# 4. Google Drive backup every 15 minutes
+
+import subprocess, threading, os, time, queue
+from datetime import datetime
+
+# Kill any existing processes first
+subprocess.run(['pkill', '-f', 'train_deberta_local'], capture_output=True)
+time.sleep(2)
+
+print("🚀 PHASE 1.5: IMPROVED PARALLEL TRAINING WITH BACKUP & MONITORING")
+print("=" * 70)
+print("IMPROVEMENTS:")
+print("✅ AsymmetricLoss fixed to maintain gradient flow")
+print("✅ Progress monitoring to auto-detect training stalls")
+print("✅ Auto-backup to Google Drive every 15 minutes")
+print("✅ Disk quota management with auto-cleanup")
+print("✅ NCCL optimizations for better multi-GPU communication")
+print("=" * 70)
+
+# RESUME COMBINED MODELS SINCE THESE WERE INTERRUPTED
+print("\n📍 RESUMING: Combined 0.7 (GPU0) + Combined 0.5 (GPU1) with BACKUP & MONITORING")
+t1 = threading.Thread(target=run_config_with_monitor_and_backup, args=(0, 'Combined_07_Parallel', False, 0.7))
+t2 = threading.Thread(target=run_config_with_monitor_and_backup, args=(1, 'Combined_05_Parallel', False, 0.5))
+t1.start()
+t2.start()
+t1.join()
+t2.join()
+
+# Run Combined 0.3 after the others are done
+print("\n📍 SINGLE: Combined 0.3 (GPU0) with BACKUP & MONITORING")
+run_config_with_monitor_and_backup(0, 'Combined_03_Parallel', False, 0.3)
+
+print("\n🎉 PHASE 1.5 PARALLEL COMPLETE with ALL IMPROVEMENTS!")
+print("📊 Outputs: ./outputs/parallel_Combined_07_Parallel/, ./outputs/parallel_Combined_05_Parallel/, etc.")
+print("💾 Backups saved to: drive:00_Projects/🎯 TechLabs-2025/Final_Project/TRAINING/GoEmotions-DeBERTa-Backup")
+print("📈 Run the analysis cell to compare F1@0.2 scores")
+print("⚠️ If training stalls, it will automatically recover or notify you")# AUTOMATIC GOOGLE DRIVE BACKUP SYSTEM
+
+# Create a backup script that automatically syncs to Google Drive
+backup_script = """#!/bin/bash
+
+# Backup script for GoEmotions-DeBERTa model outputs
+# Syncs training outputs to Google Drive to prevent disk quota issues
+
+# Target Google Drive folder
+DRIVE_TARGET="drive:00_Projects/🎯 TechLabs-2025/Final_Project/TRAINING/GoEmotions-DeBERTa-Backup"
+
+# Local paths to backup
+MODEL_OUTPUTS="./outputs"
+MODEL_CACHE="./models"
+DATASET_CACHE="./data"
+LOGS="./logs"
+
+# Timestamp
+TIMESTAMP=$(date +"%Y-%m-%d_%H-%M-%S")
+echo "🔄 Starting backup at $TIMESTAMP"
+
+# Check if rclone is configured
+if ! rclone version &>/dev/null; then
+    echo "❌ rclone not found or not configured"
+    exit 1
+fi
+
+# Check if Google Drive target exists
+if ! rclone lsd "$DRIVE_TARGET" &>/dev/null; then
+    echo "🔨 Creating backup directory..."
+    rclone mkdir "$DRIVE_TARGET"
+    rclone mkdir "$DRIVE_TARGET/outputs"
+    rclone mkdir "$DRIVE_TARGET/models"
+    rclone mkdir "$DRIVE_TARGET/data"
+    rclone mkdir "$DRIVE_TARGET/logs"
+fi
+
+# Backup completed evaluation results (smallest files, highest value)
+echo "📊 Backing up evaluation reports..."
+find "$MODEL_OUTPUTS" -name "eval_report.json" -type f | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    echo "   $file → $target_path"
+    rclone copy "$file" "$(dirname "$target_path")"
+done
+
+# Backup trained models
+echo "🤖 Backing up trained models..."
+find "$MODEL_OUTPUTS" -path "*/pytorch_model.bin" -o -path "*/model.safetensors" | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    echo "   $file → $target_path"
+    rclone copy "$file" "$(dirname "$target_path")"
+done
+
+# Backup model configs
+echo "⚙️ Backing up model configs..."
+find "$MODEL_OUTPUTS" -name "config.json" -type f | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    echo "   $file → $target_path"
+    rclone copy "$file" "$(dirname "$target_path")"
+done
+
+# Backup model cache (if it doesn't already exist)
+echo "💾 Backing up model cache..."
+rclone copy "$MODEL_CACHE" "$DRIVE_TARGET/models" --update
+
+# Backup dataset cache (if it doesn't already exist)
+echo "📊 Backing up dataset cache..."
+rclone copy "$DATASET_CACHE" "$DRIVE_TARGET/data" --update
+
+# Backup logs
+echo "📝 Backing up logs..."
+rclone copy "$LOGS" "$DRIVE_TARGET/logs" --update
+
+# Record backup in log file
+echo "$TIMESTAMP: Backup completed successfully" >> "./backup_history.log"
+rclone copy "./backup_history.log" "$DRIVE_TARGET/"
+
+# Check disk space after backup
+FREE_SPACE=$(df -h . | awk 'NR==2 {print $4}')
+USED_PERCENT=$(df -h . | awk 'NR==2 {print $5}')
+echo "💾 Disk space after backup: $FREE_SPACE free ($USED_PERCENT used)"
+
+echo "✅ Backup completed at $(date)"
+"""
+
+# Create the backup script
+with open("/home/user/goemotions-deberta/backup_to_drive.sh", "w") as f:
+    f.write(backup_script)
+!chmod +x /home/user/goemotions-deberta/backup_to_drive.sh
+
+# Create a version that's run automatically during training
+auto_backup_script = """#!/bin/bash
+
+# Auto backup script that runs during training
+# Set to backup every 15 minutes to prevent disk quota issues
+
+BACKUP_INTERVAL=900  # 15 minutes
+BACKUP_SCRIPT="/home/user/goemotions-deberta/backup_to_drive.sh"
+
+while true; do
+    # Run backup
+    echo "🔄 Running scheduled backup ($(date))"
+    bash "$BACKUP_SCRIPT"
+    
+    # Check for disk quota issues
+    FREE_SPACE=$(df -h . | awk 'NR==2 {print $4}')
+    USED_PERCENT=$(df -h . | awk 'NR==2 {print $5}')
+    USED_NUM=${USED_PERCENT%\%}
+    
+    if [ "$USED_NUM" -gt 85 ]; then
+        echo "⚠️ WARNING: High disk usage ($USED_PERCENT)"
+        echo "🧹 Cleaning old checkpoints after backup..."
+        find ./outputs -name "checkpoint-*" -type d | sort | head -n -5 | xargs rm -rf
+    fi
+    
+    echo "💤 Next backup in $(($BACKUP_INTERVAL / 60)) minutes"
+    sleep $BACKUP_INTERVAL
+done
+"""
+
+# Create the auto backup script
+with open("/home/user/goemotions-deberta/auto_backup.sh", "w") as f:
+    f.write(auto_backup_script)
+!chmod +x /home/user/goemotions-deberta/auto_backup.sh
+
+# Function to start auto backup
+def start_auto_backup():
+    """Start automatic backup to Google Drive"""
+    print("🔄 Starting automatic Google Drive backup...")
+    subprocess.Popen(
+        ["bash", "/home/user/goemotions-deberta/auto_backup.sh"], 
+        stdout=subprocess.PIPE, 
+        stderr=subprocess.STDOUT,
+        text=True,
+        bufsize=1
+    )
+    print("✅ Auto backup process started - will backup every 15 minutes")
+    print("📁 Backup target: drive:00_Projects/🎯 TechLabs-2025/Final_Project/TRAINING/GoEmotions-DeBERTa-Backup")
+
+# Run an initial backup
+print("🚀 Running initial backup to verify Google Drive connection...")
+!bash /home/user/goemotions-deberta/backup_to_drive.sh
+
+# Modify our training function to start auto backup
+def run_config_with_monitor_and_backup(gpu_id, config_name, use_asym=False, ratio=None):
+    """Run training with progress monitoring and auto backup"""
+    print(f"🚀 Starting {config_name} on GPU {gpu_id} at {datetime.now()}")
+    
+    # Start auto backup process
+    start_auto_backup()
+    
+    cmd = ['./run_with_monitor.sh', str(gpu_id), config_name]
+    
+    # Add standard args
+    cmd.extend([
+        '--model_type', 'deberta-v3-large',
+        '--per_device_train_batch_size', '4',
+        '--per_device_eval_batch_size', '8',
+        '--gradient_accumulation_steps', '4',
+        '--num_train_epochs', '2',
+        '--learning_rate', '3e-5',
+        '--lr_scheduler_type', 'cosine',
+        '--warmup_ratio', '0.15',
+        '--weight_decay', '0.01',
+        '--fp16',
+        '--max_length', '256',
+        '--max_train_samples', '20000',
+        '--max_eval_samples', '3000'
+    ])
+    
+    if use_asym:
+        cmd.append('--use_asymmetric_loss')
+    
+    if ratio is not None:
+        cmd.extend(['--use_combined_loss', '--loss_combination_ratio', str(ratio)])
+    
+    print(f"Command for {config_name}: {' '.join(cmd)}")
+    
+    # Run in subprocess with full output capture
+    process = subprocess.Popen(
+        cmd,
+        cwd='/home/user/goemotions-deberta',
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+        text=True,
+        bufsize=1
+    )
+    
+    # Read output in real-time
+    for line in iter(process.stdout.readline, ''):
+        timestamp = datetime.now().strftime('%H:%M:%S')
+        print(f"[{timestamp}] GPU {gpu_id} [{config_name}]: {line.strip()}")
+    
+    # Wait for completion and get return code
+    return_code = process.wait()
+    
+    # Run final backup after training completes
+    print(f"🔄 Running final backup for {config_name}...")
+    subprocess.run(['bash', '/home/user/goemotions-deberta/backup_to_drive.sh'], 
+                   check=False, stdout=subprocess.PIPE)
+    
+    print(f"✅ {config_name} complete on GPU {gpu_id} (return code: {return_code})")
+    return return_code
+
+print("\n✅ Google Drive backup system configured!")
+print("📋 Features:")
+print("1. Auto-backup every 15 minutes during training")
+print("2. Backs up model checkpoints, configs, and evaluation results")
+print("3. Automatically cleans up old checkpoints after backup when disk usage is high")
+print("4. Final backup runs when training completes")
+print("5. All outputs are saved to: drive:00_Projects/🎯 TechLabs-2025/Final_Project/TRAINING/GoEmotions-DeBERTa-Backup")
+print("\nRun the cell below to resume training with backup enabled!")
+
+# This cell fixes the critical issue with the AsymmetricLoss implementation
+# The problem was: disable_torch_grad_focal_loss=True created a torch.no_grad() context
+# which disconnected the gradient flow and prevented learning
+
+# Create a patched version of the training script
+!cp /workspace/notebooks/scripts/asymmetric_loss_fix.patch /home/user/goemotions-deberta/
+!cd /home/user/goemotions-deberta/ && patch -b notebooks/scripts/train_deberta_local.py < asymmetric_loss_fix.patch
+
+# Verify the fix was applied
+print("\n🔍 Verifying AsymmetricLoss fix...")
+!grep -n "disable_torch_grad_focal_loss=False" /home/user/goemotions-deberta/notebooks/scripts/train_deberta_local.py
+
+print("\n✅ Fixes applied:")
+print("1. Fixed AsymmetricLoss to maintain gradient flow (disable_torch_grad_focal_loss=False)")
+print("2. Added ProgressMonitorCallback to detect training stalls after 10 minutes")
+print("3. Added disk space monitoring and auto-cleanup")
+print("4. Optimized NCCL for better multi-GPU communication")
+print("5. Added comprehensive error handling with stack traces")
+
+print("\nThese fixes will prevent the training from stopping unexpectedly.")
+
+# Set stricter NCCL environment variables to prevent timeouts
+improved_env = """
+# Set stricter NCCL config to prevent timeouts and communication issues
+export NCCL_TIMEOUT=1800
+export NCCL_BLOCKING_WAIT=1
+export NCCL_ASYNC_ERROR_HANDLING=1
+export NCCL_SOCKET_IFNAME=lo
+export NCCL_IB_DISABLE=1
+"""
+
+# Create an improved script that includes the progress monitor
+improved_script = """
+#!/bin/bash
+
+# Usage: ./run_with_monitor.sh [GPU_ID] [CONFIG_NAME] [ADDITIONAL_ARGS...]
+GPU_ID=$1
+CONFIG_NAME=$2
+shift 2  # Remove first two args
+
+# Set environment variables
+export CUDA_VISIBLE_DEVICES=$GPU_ID
+export NCCL_TIMEOUT=1800
+export NCCL_BLOCKING_WAIT=1
+export NCCL_ASYNC_ERROR_HANDLING=1
+export NCCL_SOCKET_IFNAME=lo
+export NCCL_IB_DISABLE=1
+
+# Pre-run checks
+echo "🔍 Pre-run checks for $CONFIG_NAME on GPU $GPU_ID"
+
+# Check disk space
+FREE_SPACE=$(df -h . | awk 'NR==2 {print $4}')
+USED_PERCENT=$(df -h . | awk 'NR==2 {print $5}')
+echo "💾 Disk space: $FREE_SPACE free ($USED_PERCENT used)"
+if [[ "${USED_PERCENT%\%}" -gt 85 ]]; then
+    echo "⚠️ WARNING: Disk usage is high ($USED_PERCENT). Training may fail!"
+fi
+
+# Check if we need to clean up first
+if [[ "${USED_PERCENT%\%}" -gt 90 ]]; then
+    echo "🧹 Cleaning up old checkpoints to free space..."
+    find ./outputs -name "checkpoint-*" -type d | sort | head -n -5 | xargs rm -rf
+    echo "✅ Cleanup complete"
+    FREE_SPACE=$(df -h . | awk 'NR==2 {print $4}')
+    USED_PERCENT=$(df -h . | awk 'NR==2 {print $5}')
+    echo "💾 Disk space after cleanup: $FREE_SPACE free ($USED_PERCENT used)"
+fi
+
+# Run with stall detection via timeout
+echo "🚀 Running $CONFIG_NAME on GPU $GPU_ID with progress monitoring"
+cmd="python3 notebooks/scripts/train_deberta_local.py --output_dir ./outputs/parallel_${CONFIG_NAME} $@"
+echo "Command: $cmd"
+
+# Run the command with auto-recovery if it stalls
+MAX_ATTEMPTS=3
+ATTEMPT=1
+
+while [ $ATTEMPT -le $MAX_ATTEMPTS ]; do
+    echo "💥 Attempt $ATTEMPT of $MAX_ATTEMPTS"
+    timeout -k 60 7200 $cmd  # Kill after 2 hours (7200s) if stalled
+    EXIT_CODE=$?
+    
+    if [ $EXIT_CODE -eq 0 ]; then
+        echo "✅ $CONFIG_NAME completed successfully on GPU $GPU_ID"
+        break
+    elif [ $EXIT_CODE -eq 124 ] || [ $EXIT_CODE -eq 137 ]; then
+        echo "⚠️ $CONFIG_NAME timed out on GPU $GPU_ID (stalled)"
+        
+        # Check if we're making progress by looking at tensorboard logs
+        LAST_STEP=$(grep -o 'step [0-9]*' ./outputs/parallel_${CONFIG_NAME}/tensorboard/* 2>/dev/null | tail -1 | awk '{print $2}')
+        if [ ! -z "$LAST_STEP" ]; then
+            echo "📊 Last recorded step: $LAST_STEP"
+        fi
+        
+        # Check disk space before retrying
+        FREE_SPACE=$(df -h . | awk 'NR==2 {print $4}')
+        USED_PERCENT=$(df -h . | awk 'NR==2 {print $5}')
+        echo "💾 Disk space: $FREE_SPACE free ($USED_PERCENT used)"
+        
+        if [[ "${USED_PERCENT%\%}" -gt 90 ]]; then
+            echo "🧹 Cleaning up old checkpoints to free space..."
+            find ./outputs -name "checkpoint-*" -type d | sort | head -n -5 | xargs rm -rf
+        fi
+        
+        ATTEMPT=$((ATTEMPT+1))
+        if [ $ATTEMPT -le $MAX_ATTEMPTS ]; then
+            echo "🔄 Retrying $CONFIG_NAME on GPU $GPU_ID (attempt $ATTEMPT of $MAX_ATTEMPTS)"
+            sleep 30  # Wait before retry
+        fi
+    else
+        echo "❌ $CONFIG_NAME failed on GPU $GPU_ID with exit code $EXIT_CODE"
+        break
+    fi
+done
+
+if [ $ATTEMPT -gt $MAX_ATTEMPTS ]; then
+    echo "❌ $CONFIG_NAME failed after $MAX_ATTEMPTS attempts on GPU $GPU_ID"
+fi
+"""
+
+# Create run script with monitor and auto-recovery
+with open("/home/user/goemotions-deberta/run_with_monitor.sh", "w") as f:
+    f.write(improved_script)
+!chmod +x /home/user/goemotions-deberta/run_with_monitor.sh
+
+# NCCL environment config
+with open("/home/user/goemotions-deberta/nccl_config.sh", "w") as f:
+    f.write(improved_env)
+!chmod +x /home/user/goemotions-deberta/nccl_config.sh
+
+# Clean up disk space before starting
+!find /home/user/goemotions-deberta/outputs -name "checkpoint-*" -type d | sort | head -n -10 | xargs rm -rf 2>/dev/null
+
+import subprocess, threading, os, time, queue
+from datetime import datetime
+
+# Kill any existing processes
+subprocess.run(['pkill', '-f', 'train_deberta_local'], capture_output=True)
+time.sleep(2)
+
+print("🚀 PHASE 1.5: Parallel Dual-GPU Training - IMPROVED WITH PROGRESS MONITORING")
+print("=" * 70)
+print("IMPROVEMENTS:")
+print("✅ Progress monitoring: Auto-detects stalls and recovers")
+print("✅ Disk space checks: Monitors disk quota and cleans up if needed")
+print("✅ NCCL optimizations: Reduced timeout, localhost-only communication")
+print("✅ Error handling: Comprehensive logging with stack traces")
+print("✅ Auto-recovery: Retries on timeouts up to 3 times")
+print("=" * 70)
+
+def run_config_with_monitor(gpu_id, config_name, use_asym=False, ratio=None):
+    """Run training with progress monitoring and improved error handling"""
+    print(f"🚀 Starting {config_name} on GPU {gpu_id} at {datetime.now()}")
+    
+    cmd = ['./run_with_monitor.sh', str(gpu_id), config_name]
+    
+    # Add standard args
+    cmd.extend([
+        '--model_type', 'deberta-v3-large',
+        '--per_device_train_batch_size', '4',
+        '--per_device_eval_batch_size', '8',
+        '--gradient_accumulation_steps', '4',
+        '--num_train_epochs', '2',
+        '--learning_rate', '3e-5',
+        '--lr_scheduler_type', 'cosine',
+        '--warmup_ratio', '0.15',
+        '--weight_decay', '0.01',
+        '--fp16',
+        '--max_length', '256',
+        '--max_train_samples', '20000',
+        '--max_eval_samples', '3000'
+    ])
+    
+    if use_asym:
+        cmd.append('--use_asymmetric_loss')
+    
+    if ratio is not None:
+        cmd.extend(['--use_combined_loss', '--loss_combination_ratio', str(ratio)])
+    
+    print(f"Command for {config_name}: {' '.join(cmd)}")
+    
+    # Run in subprocess with full output capture
+    process = subprocess.Popen(
+        cmd,
+        cwd='/home/user/goemotions-deberta',
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+        text=True,
+        bufsize=1
+    )
+    
+    # Read output in real-time
+    for line in iter(process.stdout.readline, ''):
+        timestamp = datetime.now().strftime('%H:%M:%S')
+        print(f"[{timestamp}] GPU {gpu_id} [{config_name}]: {line.strip()}")
+    
+    # Wait for completion and get return code
+    return_code = process.wait()
+    print(f"✅ {config_name} complete on GPU {gpu_id} (return code: {return_code})")
+    return return_code
+
+# RESUME COMBINED MODELS SINCE THESE WERE INTERRUPTED
+print("\n📍 RESUMING: Combined 0.7 (GPU0) + Combined 0.5 (GPU1)")
+t1 = threading.Thread(target=run_config_with_monitor, args=(0, 'Combined_07_Parallel', False, 0.7))
+t2 = threading.Thread(target=run_config_with_monitor, args=(1, 'Combined_05_Parallel', False, 0.5))
+t1.start()
+t2.start()
+t1.join()
+t2.join()
+
+# Run Combined 0.3 after the others are done
+print("\n📍 SINGLE: Combined 0.3 (GPU0)")
+run_config_with_monitor(0, 'Combined_03_Parallel', False, 0.3)
+
+print("\n🎉 PHASE 1.5 PARALLEL COMPLETE with IMPROVED MONITORING!")
+print("📊 Outputs: ./outputs/parallel_Combined_07_Parallel/, ./outputs/parallel_Combined_05_Parallel/, etc.")
+print("🔍 All training output printed live above - run analysis cell for F1@0.2 comparison")
+print("⏱️ Total time: ~1.5 hours with dual-GPU parallel execution")
+print("💪 Improvements: Progress monitoring, disk quota checks, NCCL optimizations, auto-recovery"){
  "cells": [
   {
    "cell_type": "markdown",
diff --git a/notebooks/improved_backup_script.py b/notebooks/improved_backup_script.py
new file mode 100644
index 0000000..17af960
--- /dev/null
+++ b/notebooks/improved_backup_script.py
@@ -0,0 +1,255 @@
+"""
+Improved backup script for GoEmotions-DeBERTa model outputs
+Correctly handles rclone paths and directory structure
+"""
+
+backup_script = """#!/bin/bash
+
+# Backup script for GoEmotions-DeBERTa model outputs
+# Syncs training outputs to Google Drive to prevent disk quota issues
+
+# Target Google Drive folder - EXACT path from rclone config
+DRIVE_TARGET="drive:00_Projects/🎯 TechLabs-2025/Final_Project/TRAINING/GoEmotions-DeBERTa-Backup"
+
+# Local paths to backup
+MODEL_OUTPUTS="./outputs"
+MODEL_CACHE="./models"
+DATASET_CACHE="./data"
+LOGS="./logs"
+
+# Timestamp
+TIMESTAMP=$(date +"%Y-%m-%d_%H-%M-%S")
+echo "🔄 Starting backup at $TIMESTAMP"
+
+# Check if rclone is configured
+if ! rclone version &>/dev/null; then
+    echo "❌ rclone not found or not configured"
+    exit 1
+fi
+
+# Verify the remote exists and we can access it
+if ! rclone lsd "drive:" &>/dev/null; then
+    echo "❌ Cannot access 'drive:' remote. Check rclone configuration."
+    exit 1
+fi
+
+# Check/create directory structure
+echo "🔍 Verifying backup directory structure..."
+if ! rclone lsd "$DRIVE_TARGET" &>/dev/null; then
+    echo "🔨 Creating main backup directory..."
+    rclone mkdir "$DRIVE_TARGET"
+fi
+
+# Create subdirectories
+for dir in "outputs" "models" "data" "logs"; do
+    if ! rclone lsd "$DRIVE_TARGET/$dir" &>/dev/null; then
+        echo "🔨 Creating $DRIVE_TARGET/$dir"
+        rclone mkdir "$DRIVE_TARGET/$dir"
+    fi
+done
+
+# Backup eval reports first (highest value, smallest size)
+echo "📊 Backing up evaluation reports..."
+find "$MODEL_OUTPUTS" -name "eval_report.json" -type f | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    target_dir=$(dirname "$target_path")
+    
+    # Ensure parent directory exists
+    rclone mkdir "$target_dir" 2>/dev/null
+    
+    echo "   $file → $target_path"
+    rclone copy "$file" "$target_dir" --progress
+done
+
+# Backup model weights (most important for resuming)
+echo "🤖 Backing up model weights..."
+find "$MODEL_OUTPUTS" -name "pytorch_model.bin" -o -name "model.safetensors" | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    target_dir=$(dirname "$target_path")
+    
+    # Ensure parent directory exists
+    rclone mkdir "$target_dir" 2>/dev/null
+    
+    echo "   $file → $target_path"
+    rclone copy "$file" "$target_dir" --progress
+done
+
+# Backup model configs
+echo "⚙️ Backing up model configs..."
+find "$MODEL_OUTPUTS" -name "config.json" -o -name "special_tokens_map.json" -o -name "tokenizer_config.json" -type f | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    target_dir=$(dirname "$target_path")
+    
+    # Ensure parent directory exists
+    rclone mkdir "$target_dir" 2>/dev/null
+    
+    echo "   $file → $target_path"
+    rclone copy "$file" "$target_dir"
+done
+
+# Backup base model cache (smaller files first)
+echo "💾 Backing up model metadata..."
+find "$MODEL_CACHE" -name "*.json" -type f | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    target_dir=$(dirname "$target_path")
+    
+    rclone mkdir "$target_dir" 2>/dev/null
+    rclone copy "$file" "$target_dir"
+done
+
+# Backup tokenizer files
+echo "🔤 Backing up tokenizer files..."
+find "$MODEL_CACHE" -name "*.model" -o -name "*.txt" -type f | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    target_dir=$(dirname "$target_path")
+    
+    rclone mkdir "$target_dir" 2>/dev/null
+    rclone copy "$file" "$target_dir"
+done
+
+# Backup dataset files (important but can be large)
+echo "📊 Backing up dataset metadata..."
+find "$DATASET_CACHE" -name "*.json" -type f | while read -r file; do
+    rel_path=$(echo "$file" | sed "s|^\./||")
+    target_path="$DRIVE_TARGET/$rel_path"
+    target_dir=$(dirname "$target_path")
+    
+    rclone mkdir "$target_dir" 2>/dev/null
+    rclone copy "$file" "$target_dir"
+done
+
+# Backup logs
+echo "📝 Backing up logs..."
+rclone copy "$LOGS" "$DRIVE_TARGET/logs" --update
+
+# Record backup in log file
+echo "$TIMESTAMP: Backup completed successfully" >> "./backup_history.log"
+rclone copy "./backup_history.log" "$DRIVE_TARGET/"
+
+# Check disk space after backup
+FREE_SPACE=$(df -h . | awk 'NR==2 {print $4}')
+USED_PERCENT=$(df -h . | awk 'NR==2 {print $5}')
+echo "💾 Disk space after backup: $FREE_SPACE free ($USED_PERCENT used)"
+
+echo "✅ Backup completed at $(date)"
+"""
+
+# Auto backup script that runs during training
+auto_backup_script = """#!/bin/bash
+
+# Auto backup script that runs during training
+# Set to backup every 15 minutes to prevent disk quota issues
+
+BACKUP_INTERVAL=900  # 15 minutes
+BACKUP_SCRIPT="/home/user/goemotions-deberta/backup_to_drive.sh"
+
+echo "🔄 Starting automatic backup service at $(date)"
+echo "📁 Target: drive:00_Projects/🎯 TechLabs-2025/Final_Project/TRAINING/GoEmotions-DeBERTa-Backup"
+echo "⏱️ Backup interval: $BACKUP_INTERVAL seconds (15 minutes)"
+echo "🧹 Auto-cleanup: Enabled for disk usage >85%"
+
+while true; do
+    # Run backup
+    echo ""
+    echo "🔄 Running scheduled backup ($(date))"
+    bash "$BACKUP_SCRIPT"
+    
+    # Check for disk quota issues
+    FREE_SPACE=$(df -h . | awk 'NR==2 {print $4}')
+    USED_PERCENT=$(df -h . | awk 'NR==2 {print $5}')
+    USED_NUM=${USED_PERCENT%\%}
+    
+    if [ "$USED_NUM" -gt 85 ]; then
+        echo "⚠️ WARNING: High disk usage ($USED_PERCENT)"
+        echo "🧹 Cleaning old checkpoints after backup..."
+        # Find and list what will be removed first
+        echo "Finding old checkpoints to remove..."
+        find ./outputs -name "checkpoint-*" -type d | sort | head -n -5
+        # Then remove them
+        find ./outputs -name "checkpoint-*" -type d | sort | head -n -5 | xargs rm -rf
+        
+        # Check space after cleanup
+        FREE_SPACE=$(df -h . | awk 'NR==2 {print $4}')
+        USED_PERCENT=$(df -h . | awk 'NR==2 {print $5}')
+        echo "💾 Disk space after cleanup: $FREE_SPACE free ($USED_PERCENT used)"
+    fi
+    
+    echo "💤 Next backup in $(($BACKUP_INTERVAL / 60)) minutes ($(date -d "+$BACKUP_INTERVAL seconds"))"
+    sleep $BACKUP_INTERVAL
+done
+"""
+
+# Function to test rclone connection
+test_rclone_script = """#!/bin/bash
+
+# Test rclone connection to Google Drive
+echo "🔍 Testing rclone connection to Google Drive..."
+
+# Check if rclone is installed
+if ! command -v rclone &> /dev/null; then
+    echo "❌ rclone not found. Please install it first."
+    exit 1
+fi
+
+# Check if drive remote exists
+if ! rclone listremotes | grep -q "drive:"; then
+    echo "❌ 'drive:' remote not found in rclone configuration."
+    echo "Please configure rclone with 'rclone config' command first."
+    exit 1
+fi
+
+# Try to list root directory
+echo "📁 Testing access to drive: remote..."
+if ! rclone lsd drive: &>/dev/null; then
+    echo "❌ Cannot access 'drive:' remote. Please check your rclone configuration."
+    exit 1
+fi
+
+# Try to access or create the backup directory
+TARGET_DIR="drive:00_Projects/🎯 TechLabs-2025/Final_Project/TRAINING/GoEmotions-DeBERTa-Backup"
+echo "📁 Testing access to target directory: $TARGET_DIR"
+
+if ! rclone lsd "$TARGET_DIR" &>/dev/null; then
+    echo "🔨 Target directory doesn't exist. Creating it now..."
+    if ! rclone mkdir "$TARGET_DIR" &>/dev/null; then
+        echo "❌ Failed to create target directory. Please check permissions and path."
+        exit 1
+    else
+        echo "✅ Target directory created successfully."
+    fi
+else
+    echo "✅ Target directory exists and is accessible."
+fi
+
+# Try writing a test file
+TEST_FILE="/tmp/rclone_test_$(date +%s).txt"
+echo "This is a test file. Created at $(date)" > "$TEST_FILE"
+
+echo "📤 Uploading test file to $TARGET_DIR..."
+if ! rclone copy "$TEST_FILE" "$TARGET_DIR" &>/dev/null; then
+    echo "❌ Failed to upload test file. Please check permissions."
+    rm -f "$TEST_FILE"
+    exit 1
+else
+    echo "✅ Test file uploaded successfully."
+    
+    # Try to read the file back
+    echo "📥 Verifying test file..."
+    if ! rclone ls "$TARGET_DIR/$(basename "$TEST_FILE")" &>/dev/null; then
+        echo "⚠️ Warning: File uploaded but not immediately visible. This may be normal for Google Drive."
+    else
+        echo "✅ File verified in target location."
+        # Clean up the test file
+        rclone delete "$TARGET_DIR/$(basename "$TEST_FILE")" &>/dev/null
+    fi
+fi
+
+rm -f "$TEST_FILE"
+echo "✅ rclone connection test completed successfully!"
+echo "✅ Backup system ready to use with target: $TARGET_DIR"
+"""
\ No newline at end of file
diff --git a/notebooks/scripts/asymmetric_loss_fix.patch b/notebooks/scripts/asymmetric_loss_fix.patch
new file mode 100644
index 0000000..3bff3ea
--- /dev/null
+++ b/notebooks/scripts/asymmetric_loss_fix.patch
@@ -0,0 +1,250 @@
+--- train_deberta_local.py.orig	2025-09-09 15:00:00.000000000 +0000
++++ train_deberta_local.py	2025-09-09 15:10:00.000000000 +0000
+@@ -26,6 +26,9 @@
+ os.environ["NCCL_TIMEOUT"] = "1800"  # 30 minutes timeout (reduced from 1 hour)
+ os.environ["NCCL_BLOCKING_WAIT"] = "1"  # Enable blocking wait
+ os.environ["NCCL_ASYNC_ERROR_HANDLING"] = "1"  # Better error handling
++os.environ["NCCL_SOCKET_IFNAME"] = "lo"  # Force localhost-only communication
++os.environ["NCCL_IB_DISABLE"] = "1"  # Disable InfiniBand
++os.environ["NCCL_DEBUG"] = "INFO"  # Enable debug info for troubleshooting
+ 
+ from typing import List, Dict, Any
+ import torch
+@@ -35,7 +38,8 @@
+ from torch.utils.data import Dataset
+ from transformers import (
+     AutoTokenizer, AutoModelForSequenceClassification, AutoConfig,
+-    Trainer, TrainingArguments, DataCollatorWithPadding
++    Trainer, TrainingArguments, DataCollatorWithPadding,
++    TrainerCallback, TrainerState, TrainerControl
+ )
+ from sklearn.metrics import f1_score, precision_recall_fscore_support
+ import logging
+@@ -43,6 +47,9 @@
+ import random
+ from datetime import datetime
+ 
++import sys
++import traceback
++import shutil
+ 
+ # Set up logging
+ logging.basicConfig(level=logging.INFO)
+@@ -123,6 +130,67 @@
+         with open(self.log_file, 'a') as f:
+             f.write(json.dumps(log_data) + '\n')
+ 
++# ProgressMonitorCallback to detect training stalls
++class ProgressMonitorCallback(TrainerCallback):
++    """Monitors progress and detects stalls in training"""
++    
++    def __init__(self, stall_timeout=600, disk_quota_check=True, min_free_space_gb=10):
++        """Initialize ProgressMonitor
++        
++        Args:
++            stall_timeout: Seconds without progress before considering training stalled
++            disk_quota_check: Whether to check disk space
++            min_free_space_gb: Minimum free disk space in GB before warning
++        """
++        self.last_step = 0
++        self.last_progress_time = time.time()
++        self.stall_timeout = stall_timeout
++        self.disk_quota_check = disk_quota_check
++        self.min_free_space_gb = min_free_space_gb
++        self.check_disk_space()
++    
++    def check_disk_space(self):
++        """Check available disk space"""
++        if not self.disk_quota_check:
++            return True
++        
++        try:
++            # Get disk usage of current directory
++            disk = shutil.disk_usage('.')
++            free_gb = disk.free / (1024 ** 3)
++            total_gb = disk.total / (1024 ** 3)
++            used_percent = (disk.used / disk.total) * 100
++            
++            print(f"💾 Disk space: {free_gb:.1f}GB free / {total_gb:.1f}GB total ({used_percent:.1f}% used)")
++            
++            if free_gb < self.min_free_space_gb:
++                print(f"⚠️ LOW DISK SPACE WARNING: Only {free_gb:.1f}GB free. Training may fail with disk quota error.")
++                if used_percent > 85:
++                    print("🔥 CRITICAL: Disk usage above 85%. Try to free up space immediately!")
++                return False
++                
++            return True
++        except Exception as e:
++            print(f"⚠️ Error checking disk space: {e}")
++            return True
++    
++    def on_step_end(self, args, state, control, **kwargs):
++        """Called after each step"""
++        # Update progress tracker if step increased
++        if state.global_step > self.last_step:
++            self.last_step = state.global_step
++            self.last_progress_time = time.time()
++        
++        # Check if we're stalled
++        time_since_progress = time.time() - self.last_progress_time
++        if time_since_progress > self.stall_timeout:
++            print(f"⚠️ WARNING: No progress for {time_since_progress:.1f} seconds (stall detected)")
++            print(f"🔍 Last progress at step {self.last_step}")
++            
++            # Check disk space on stall detection
++            if not self.check_disk_space():
++                print("❌ TRAINING STOPPED: Disk space critical. Clean up files before continuing.")
++                raise RuntimeError("Training stopped due to disk quota issue")
++        
++        return control
+ 
+ # GoEmotions labels
+ EMOTION_LABELS = [
+@@ -139,7 +207,7 @@
+-    def __init__(self, gamma_neg=1.0, gamma_pos=1.0, clip=0.2, eps=1e-8, disable_torch_grad_focal_loss=True):
++    def __init__(self, gamma_neg=1.0, gamma_pos=1.0, clip=0.2, eps=1e-8, disable_torch_grad_focal_loss=False):
+         super(AsymmetricLoss, self).__init__()
+         self.gamma_neg = gamma_neg
+         self.gamma_pos = gamma_pos
+@@ -173,12 +241,7 @@
+         if self.gamma_neg > 0 or self.gamma_pos > 0:
+             # FIXED: Remove no_grad for full differentiability (HF docs compliant)
+             if self.disable_torch_grad_focal_loss:
+-                with torch.no_grad():
+-                    pt0 = xs_pos * y
+-                    pt1 = xs_neg * (1 - y)  # pt = p if t > 0 else 1-p
+-                    pt = pt0 + pt1
+-                    one_sided_gamma = self.gamma_pos * y + self.gamma_neg * (1 - y)
+-                    one_sided_w = torch.pow(1 - pt, one_sided_gamma)
++                # WARNING: This code path has gradient disconnection - never use this!
+                 loss = loss * one_sided_w
+             else:
+                 pt0 = xs_pos * y
+@@ -247,7 +310,7 @@
+     """
+     def __init__(self, loss_combination_ratio=0.7, *args, **kwargs):
+         super().__init__(*args, **kwargs)
+-        self.asymmetric_loss = AsymmetricLoss(gamma_neg=1.0, gamma_pos=1.0, clip=0.2, disable_torch_grad_focal_loss=True)
++        self.asymmetric_loss = AsymmetricLoss(gamma_neg=1.0, gamma_pos=1.0, clip=0.2, disable_torch_grad_focal_loss=False)
+         self.focal_loss = FocalLoss(alpha=0.25, gamma=2.0)
+         self.loss_combination_ratio = loss_combination_ratio
+ 
+@@ -384,7 +447,7 @@
+     """
+     def __init__(self, *args, **kwargs):
+         super().__init__(*args, **kwargs)
+-        self.asymmetric_loss = AsymmetricLoss(gamma_neg=1.0, gamma_pos=1.0, clip=0.2, disable_torch_grad_focal_loss=True)
++        self.asymmetric_loss = AsymmetricLoss(gamma_neg=1.0, gamma_pos=1.0, clip=0.2, disable_torch_grad_focal_loss=False)
+ 
+     def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
+         """
+@@ -680,6 +743,42 @@
+         return None, None
+ 
+ 
++def check_system_status():
++    """Check system status and resources"""
++    print("\n🔍 Checking system status...")
++    
++    # Check disk space
++    try:
++        disk = shutil.disk_usage('.')
++        free_gb = disk.free / (1024 ** 3)
++        total_gb = disk.total / (1024 ** 3)
++        used_percent = (disk.used / disk.total) * 100
++        print(f"💾 Disk space: {free_gb:.1f}GB free / {total_gb:.1f}GB total ({used_percent:.1f}% used)")
++        
++        if free_gb < 10:
++            print(f"⚠️ LOW DISK SPACE WARNING: Only {free_gb:.1f}GB free")
++        if used_percent > 85:
++            print("🔥 CRITICAL: Disk usage above 85%. Training may fail!")
++    except Exception as e:
++        print(f"⚠️ Error checking disk space: {e}")
++    
++    # Check GPU status
++    try:
++        if torch.cuda.is_available():
++            for i in range(torch.cuda.device_count()):
++                print(f"GPU {i}: {torch.cuda.get_device_name(i)}")
++                print(f"   Memory: {torch.cuda.memory_allocated(i)/1024**2:.1f}MB allocated, "
++                      f"{torch.cuda.memory_reserved(i)/1024**2:.1f}MB reserved")
++        else:
++            print("❌ No GPUs available")
++    except Exception as e:
++        print(f"⚠️ Error checking GPU status: {e}")
++    
++    print("✅ System check complete\n")
++
++
++def handle_exception(e, scientific_logger=None):
++    """Handle exceptions with comprehensive error logging"""
++    error_msg = f"❌ ERROR: {str(e)}"
++    print(error_msg)
++    print("Stack trace:")
++    traceback_str = traceback.format_exc()
++    print(traceback_str)
++    
++    # Log to scientific logger if available
++    if scientific_logger:
++        error_log = {
++            "timestamp": datetime.now().isoformat(),
++            "error": str(e),
++            "traceback": traceback_str,
++        }
++        scientific_logger._write_log(error_log)
++    
++    return 1
++
++
+ def main():
+     parser = argparse.ArgumentParser()
+     parser.add_argument("--output_dir", type=str, default="./outputs/deberta")
+@@ -717,10 +816,13 @@
+     print(f"📊 Dataset: GoEmotions (from local cache)")
+     print(f"🔬 Scientific logging: ENABLED")
+     
++    # Check system resources
++    check_system_status()
++    
+     # Create output directory
+-    os.makedirs(args.output_dir, exist_ok=True)
++    try:
++        os.makedirs(args.output_dir, exist_ok=True)
++    except Exception as e:
++        print(f"❌ ERROR: Could not create output directory: {e}")
++        print("💡 TIP: Check disk quota or remove old checkpoints")
++        return 1
+     
+     # Initialize scientific logger
+     scientific_logger = ScientificLogger(args.output_dir)
+@@ -813,6 +915,9 @@
+             compute_metrics=compute_comprehensive_metrics,
+         )
+     
++    # Add progress monitoring callback
++    progress_monitor = ProgressMonitorCallback(stall_timeout=600)  # 10 minutes timeout
++    trainer.add_callback(progress_monitor)
+     
+     # Train
+     print("🚀 Starting training...")
+-    trainer.train()
++    try:
++        trainer.train()
++    except Exception as e:
++        handle_exception(e, scientific_logger)
++        return 1
+     
+     # Save final model
+     trainer.save_model()
+@@ -904,4 +1009,10 @@
+ 
+ 
+ if __name__ == "__main__":
+-    main()
++    try:
++        main()
++    except Exception as e:
++        print(f"❌ UNHANDLED ERROR: {str(e)}")
++        traceback.print_exc()
++        sys.exit(1)
++
\ No newline at end of file
diff --git a/notebooks/scripts/progress_monitor_callback.py b/notebooks/scripts/progress_monitor_callback.py
new file mode 100644
index 0000000..4827f15
--- /dev/null
+++ b/notebooks/scripts/progress_monitor_callback.py
@@ -0,0 +1,135 @@
+"""
+Progress Monitor Callback for Transformers training
+Detects stalls in training and checks disk space
+"""
+
+import time
+import shutil
+import traceback
+from datetime import datetime
+from transformers import TrainerCallback, TrainerState, TrainerControl
+
+class ProgressMonitorCallback(TrainerCallback):
+    """Monitors progress and detects stalls in training"""
+    
+    def __init__(self, stall_timeout=600, disk_quota_check=True, min_free_space_gb=10):
+        """Initialize ProgressMonitor
+        
+        Args:
+            stall_timeout: Seconds without progress before considering training stalled
+            disk_quota_check: Whether to check disk space
+            min_free_space_gb: Minimum free disk space in GB before warning
+        """
+        self.last_step = 0
+        self.last_progress_time = time.time()
+        self.stall_timeout = stall_timeout
+        self.disk_quota_check = disk_quota_check
+        self.min_free_space_gb = min_free_space_gb
+        self.check_disk_space()
+    
+    def check_disk_space(self):
+        """Check available disk space"""
+        if not self.disk_quota_check:
+            return True
+        
+        try:
+            # Get disk usage of current directory
+            disk = shutil.disk_usage('.')
+            free_gb = disk.free / (1024 ** 3)
+            total_gb = disk.total / (1024 ** 3)
+            used_percent = (disk.used / disk.total) * 100
+            
+            print(f"💾 Disk space: {free_gb:.1f}GB free / {total_gb:.1f}GB total ({used_percent:.1f}% used)")
+            
+            if free_gb < self.min_free_space_gb:
+                print(f"⚠️ LOW DISK SPACE WARNING: Only {free_gb:.1f}GB free. Training may fail with disk quota error.")
+                if used_percent > 85:
+                    print("🔥 CRITICAL: Disk usage above 85%. Try to free up space immediately!")
+                return False
+                
+            return True
+        except Exception as e:
+            print(f"⚠️ Error checking disk space: {e}")
+            return True
+    
+    def on_step_end(self, args, state, control, **kwargs):
+        """Called after each step"""
+        # Update progress tracker if step increased
+        if state.global_step > self.last_step:
+            self.last_step = state.global_step
+            self.last_progress_time = time.time()
+        
+        # Check if we're stalled
+        time_since_progress = time.time() - self.last_progress_time
+        if time_since_progress > self.stall_timeout:
+            print(f"⚠️ WARNING: No progress for {time_since_progress:.1f} seconds (stall detected)")
+            print(f"🔍 Last progress at step {self.last_step}")
+            
+            # Check disk space on stall detection
+            if not self.check_disk_space():
+                print("❌ TRAINING STOPPED: Disk space critical. Clean up files before continuing.")
+                # We cannot set control.should_training_stop because it's read-only
+                # Instead we'll raise an exception to stop training
+                raise RuntimeError("Training stopped due to disk quota issue")
+        
+        # Check disk space every 100 steps
+        if state.global_step % 100 == 0 and self.disk_quota_check:
+            self.check_disk_space()
+        
+        return control
+        
+def check_system_status():
+    """Check system status and resources"""
+    print("\n🔍 Checking system status...")
+    
+    # Check disk space
+    try:
+        disk = shutil.disk_usage('.')
+        free_gb = disk.free / (1024 ** 3)
+        total_gb = disk.total / (1024 ** 3)
+        used_percent = (disk.used / disk.total) * 100
+        print(f"💾 Disk space: {free_gb:.1f}GB free / {total_gb:.1f}GB total ({used_percent:.1f}% used)")
+        
+        if free_gb < 10:
+            print(f"⚠️ LOW DISK SPACE WARNING: Only {free_gb:.1f}GB free")
+        if used_percent > 85:
+            print("🔥 CRITICAL: Disk usage above 85%. Training may fail!")
+    except Exception as e:
+        print(f"⚠️ Error checking disk space: {e}")
+    
+    # Check GPU status
+    try:
+        import torch
+        if torch.cuda.is_available():
+            for i in range(torch.cuda.device_count()):
+                print(f"GPU {i}: {torch.cuda.get_device_name(i)}")
+                print(f"   Memory: {torch.cuda.memory_allocated(i)/1024**2:.1f}MB allocated, "
+                      f"{torch.cuda.memory_reserved(i)/1024**2:.1f}MB reserved")
+        else:
+            print("❌ No GPUs available")
+    except Exception as e:
+        print(f"⚠️ Error checking GPU status: {e}")
+    
+    print("✅ System check complete\n")
+
+def handle_exception(e, scientific_logger=None):
+    """Handle exceptions with comprehensive error logging"""
+    error_msg = f"❌ ERROR: {str(e)}"
+    print(error_msg)
+    print("Stack trace:")
+    traceback_str = traceback.format_exc()
+    print(traceback_str)
+    
+    # Log to scientific logger if available
+    if scientific_logger:
+        error_log = {
+            "timestamp": datetime.now().isoformat(),
+            "error": str(e),
+            "traceback": traceback_str,
+        }
+        try:
+            scientific_logger._write_log(error_log)
+        except:
+            pass
+    
+    return 1
\ No newline at end of file