projects/elizabeth/training/start_training_mlflow.sh

#!/bin/bash

# Elizabeth Training Launcher with MLflow
# Complete experiment tracking and versioning

set -e

# Configuration
MODEL_DIR="/home/x/adaptai/experiments/qwen3-8b-elizabeth-sft"
LOG_DIR="/home/x/adaptai/training_logs"
TENSORBOARD_DIR="/home/x/adaptai/tensorboard"
MLFLOW_DIR="/home/x/adaptai/mlflow"
EXPERIMENT_DIR="/home/x/adaptai/experiments"

# Create directories
mkdir -p "$MODEL_DIR" "$LOG_DIR" "$TENSORBOARD_DIR" "$MLFLOW_DIR" "$EXPERIMENT_DIR"

# Start MLflow server in background
echo "🔬 Starting MLflow server..."
nohup mlflow server \
    --backend-store-uri "file://$MLFLOW_DIR" \
    --default-artifact-root "$EXPERIMENT_DIR" \
    --host 0.0.0.0 \
    --port 5000 \
    > "$LOG_DIR/mlflow_server.log" 2>&1 &
MLFLOW_PID=$!
echo "MLflow server started with PID: $MLFLOW_PID"

# Start TensorBoard in background
echo "📊 Starting TensorBoard..."
nohup tensorboard --logdir="$TENSORBOARD_DIR" --port=6006 --bind_all > "$LOG_DIR/tensorboard.log" 2>&1 &
TENSORBOARD_PID=$!
echo "TensorBoard started with PID: $TENSORBOARD_PID"

# Training command
echo "🚀 Starting Elizabeth training with full tracking..."
echo "⏰ Start time: $(date -Iseconds)"
echo "📈 MLflow UI: http://$(hostname -I | awk '{print $1}'):5000"
echo "📊 TensorBoard: http://$(hostname -I | awk '{print $1}'):6006"

# Run training with MLflow tracking
python3 /home/x/adaptai/aiml/train_elizabeth_mlflow.py \
    --output_dir "$MODEL_DIR" \
    --logging_dir "$TENSORBOARD_DIR" \
    2>&1 | tee "$LOG_DIR/training_$(date +%Y%m%d_%H%M%S).log"

# Save training completion time
echo "✅ Training completed!"
echo "⏰ End time: $(date -Iseconds)"

# Stop services
kill $MLFLOW_PID
kill $TENSORBOARD_PID
echo "Services stopped"

# Save comprehensive training info
echo "💾 Saving comprehensive training information..."
cat > "$MODEL_DIR/training_info.md" << EOF
# Elizabeth Nova Training - Complete Run

## Experiment Details
- **Start Time**: $(date -Iseconds)
- **End Time**: $(date -Iseconds)
- **Model**: Qwen3-8B with YaRN 128K
- **GPUs**: 2x H200
- **Training Examples**: 2879
- **Validation Examples**: 319
- **Context Length**: 131072
- **Batch Size**: 32 (effective)
- **Learning Rate**: 1.2e-5
- **Epochs**: 2

## Tracking & Monitoring
- **MLflow Server**: http://$(hostname -I | awk '{print $1}'):5000
- **TensorBoard**: http://$(hostname -I | awk '{print $1}'):6006
- **Log Directory**: $LOG_DIR
- **Model Directory**: $MODEL_DIR
- **MLflow Storage**: $MLFLOW_DIR

## Dataset Composition
- Tool Use Examples: 198
- Synthetic Corpus: 2681
- Total Training: 2879
- Validation: 319 (10%)

## Architecture Features
- Pure weight drift (no LoRA/adapters)
- YaRN RoPE scaling for 128K context
- BFloat16 precision
- Gradient checkpointing
- Tensor parallelism on 2x H200

## Versioning & Lineage
- All logs preserved permanently
- MLflow experiment tracking
- Model registry with versioning
- Complete dataset lineage

## Next Steps
1. Evaluate tool call accuracy
2. Deploy to vLLM serving
3. Set up continuous evaluation
4. Begin autonomous training cycle
EOF

echo "🎉 Training finished with complete tracking!"
echo "📈 MLflow experiments: $MLFLOW_DIR"
echo "📊 TensorBoard logs: $TENSORBOARD_DIR"
echo "💾 Model artifacts: $MODEL_DIR"
echo "📝 Training logs: $LOG_DIR"
echo "🔗 MLflow UI: http://$(hostname -I | awk '{print $1}'):5000"