Fine-Tuning Guide
This step-by-step tutorial walks through configuring and launching a training job.Prerequisites
- Workspace with at least one dataset.
- Role permission:
fine-tune.launch. - Understanding of dataset structure (see Datasets).
Step 1: Choose Base Model
- Open Fine-Tune Studio.
- Review recommended models. The panel shows context length, latency expectations, and suitable strategies.
- Select the foundation model. Record the model ID for future reference.
Step 2: Dataset Options
- Select Uploaded Dataset or connect external sources.
- Optionally blend multiple datasets:
- Assign weights (e.g.,
support_chat:0.7,policy_docs:0.3). - Preview the combined distribution.
- Assign weights (e.g.,
- Confirm the final dataset configuration.
Step 3: Configure Training
Key settings:| Setting | Description |
|---|---|
| Training method | LoRA, QLoRA, or full fine-tune |
| Learning rate | Default is 2e-4 for LoRA; adjust for specific datasets |
| Batch size | Adaptive; suggested ranges appear based on tokens per row |
| Epochs | Typically 2–5 for LoRA; validate with evaluation metrics |
| Evaluation cadence | Frequency of validation runs |
| Safety guardrails | Enable automatic redaction and safety classifiers |
| Logging integrations | Optional: send metrics to Slack, DataDog, or webhook |
Step 4: Launch and Monitor
- Confirm the cost estimate.
- Click Launch Training.
- Monitor progress in real time:
- Loss curves
- Tokens/sec
- Epoch completion
- GPU allocation events
- You can pause or cancel if necessary.
Step 5: Review Results
When the run completes:- Inspect evaluation metrics and sample outputs.
- Use the Evaluate feature to test your model:
- Single prompt evaluation
- Comparison with base model
- Batch evaluation for multiple prompts
- Results dashboard for comprehensive analysis
- Compare with previous runs using the comparison view.
- Archive the run or mark as ready for deployment.
Troubleshooting
- Job stuck in pending: Check resource availability or contact support if it exceeds expected queue times.
- Spiking loss: Inspect dataset balance; consider re-weighting or cleaning data.
- Unexpected outputs: Revisit dataset quality and ensure conversation turns are labeled correctly.