Jake AI Training Steps

1. Start Compute (2 minutes)

Go to Azure ML Studio
Compute > Compute Instances
Start jake-ai-training instance
Wait for status "Running"

2. Upload Data (3 minutes)

Click "Data" in sidebar
Create new datastore
Upload jake_training.jsonl
Name it "jake-training-data"

3. Start Training (5 minutes)

Open JupyterLab
Upload training files
Run training script
Verify training started

4. Monitor Training (2-3 hours)

Watch training metrics
Check loss values
Monitor GPU usage

5. Save Model

Training auto-saves checkpoints
Final model saved automatically
Copy to your storage

Latest Training (December 4, 2024)

Successfully fine-tuned Yi-34B model using Azure ML A100 GPU
Achieved loss reduction from 4.1917 to 2.069 over 3 epochs
Used 4-bit quantization and LoRA for efficient training
Training Parameters:
- Batch size: 1
- Gradient accumulation steps: 16
- Learning rate: 0.0001
- Epochs: 3
- QLoRA configuration:
  - LoRA rank: 64
  - LoRA alpha: 128
  - Target modules: ["q_proj", "k_proj", "v_proj", "o_proj"]
  - 4-bit quantization (bfloat16)

Model Artifacts

Platform: Azure ML
Environment: azureml_py38
GPU: A100
Cache: /dev/shm/cache

Testing

Created test_model.py for inference with cleaned prompt templates
Focused on capturing Jake's unique sales storytelling style
Implemented comprehensive data cleaning pipeline

Important Notes

Estimated cost: ~$10.80
Training time: 2-3 hours
Auto-shutdown enabled
Checkpoints every 50 steps

Troubleshooting

If training stops:

Check GPU memory usage
Verify data loading
Check training logs
Restart from checkpoint if needed

Cost Saving Tips

Start instance right before training
Use auto-shutdown
Monitor training progress
Stop instance if not needed

Success Metrics

Loss should decrease steadily
Check generated text quality
Save best checkpoint

Remember: You can monitor training remotely - no need to keep your computer on!