The Fine-Tuning Pipeline
Click each stage to see what happens there.
Select a stage above to learn about it.
Training Loss Simulator
Adjust hyperparameters and run a simulated fine-tuning job. Watch training loss (teal) versus validation loss (violet). Too many epochs or too little data causes overfitting: validation loss climbs while training loss keeps falling.
Press Run training to start.
Full Fine-Tuning vs LoRA vs Prompting
Interviewers often ask when to fine-tune versus adapt more cheaply. Here is the honest trade-off table.
| Method | Params updated | Data needed | Cost | Best for |
|---|---|---|---|---|
| Full fine-tuning | All weights | 10k+ labeled examples | High (multi-GPU) | Deep domain shift, new behavior |
| LoRA / PEFT | ~0.1–1% (low-rank adapters) | 1k–50k examples | Low (single GPU) | Task specialization on a budget |
| Prompting / RAG | None | Few examples or documents | Minimal | Fast iteration, fresh knowledge |
Rule of thumb: start with prompting, add retrieval for knowledge, and fine-tune when you need consistent style, format, or task behavior that prompts cannot reliably enforce.