fine-tuning

# Fine-Tuning ## Purpose Decide whether fine-tuning is warranted, and do it properly if it is. Most fine-tuning projects should have been prompt engineering or retrieval, and the ones that should be fine-tuning usually fail on dataset quality rather than on the training. ## When to Use - A task where prompting has plateaued below the required accuracy. - A specific output format, style, or domain vocabulary the model will not adopt reliably. - Reducing cost by making a small model do what currently requires a large one. - Evaluating an existing fine-tuned model that underperforms. ## Capabilities - Deciding between prompting, RAG, and fine-tuning. - Dataset construction, curation, and splitting. - LoRA and QLoRA versus full fine-tuning. - Hyperparameter selection and overfitting detection. - Evaluation against the base model on the same set. ## Inputs - The task, and the accuracy prompting achieves on it. - Available training data — its volume, and honestly, its quality. - Latency and cost constraints. ## Outputs - A justified decision to fine-tune, or not to. - A curated dataset with clean train/validation/test splits. - A model measurably better than the base model on a held-out set. ## Workflow 1. **Exhaust prompting first** — Few-shot examples, a clearer output contract, a better model. Fine-tuning cannot teach knowledge the model lacks; it teaches behavior. If the problem is that the model does not *know* something, use retrieval instead. 2. **Decide what fin...