Learning path · Models & Tokens · 15
Fine-Tuning
Continued training on curated data to adapt model behaviour—tone, format, domain language—without replacing base reasoning capabilities.
Why it matters
- Useful when prompt engineering plateaus on consistent style or classification.
- Not a substitute for RAG when facts change frequently.
- Requires data governance, eval splits, and regression monitoring.
Key ideas
- Behaviour adaptation
- SFT and preference tuning
- Data quality
Fine-tuning updates weights to absorb repetitive patterns: support macros, internal taxonomy, formatting habits. It is the wrong first lever for dynamic knowledge—retrieve instead. Successful fine-tunes start with a narrow objective, hundreds to thousands of high-quality examples, and held-out evals that mirror production failures. Treat fine-tuned models like any dependency: version datasets, track eval metrics, and plan rollback when base models improve. Maintain a canonical eval set of production failures before each fine-tune; if the new checkpoint wins on style but loses on facts, reject it.
Updated 2026-08-09 · Full learning path