Fine-Tuning & Custom Models
When off-the-shelf LLMs aren't enough: fine-tune models or use RAG for domain-specific knowledge.
Approaches (Easiest → Hardest)
- RAG (Retrieval Augmented Generation): Add context via vector DB. No training needed. (Easiest, recommended)
- Fine-Tuning: Adapt existing model on your data. Requires 100s-1000s examples. (Intermediate)
- Pre-Training: Train model from scratch. Requires millions of examples, huge compute. (Hardest, rarely needed)
When to Fine-Tune
- • Company has unique coding standards
- • Working with legacy frameworks LLMs don't know
- • Domain-specific language (e.g., internal DSL)
- • Need to reduce latency (smaller fine-tuned model)
Tools
- OpenAI Fine-Tuning: GPT-3.5/4 fine-tuning API
- Anthropic: Coming soon for Claude
- Open-source: Llama 2/3, Mistral (run locally)