Fine-tuning

Fine-tuning is the process of further training a base LLM on your own data so it learns patterns, vocabulary, or formats specific to your use case.

Fine-tuning bakes behavior into the model weights. It's powerful for stable patterns: brand voice, structured output formats, domain vocabulary. It's the wrong tool for changing facts (use RAG) or for tasks where the prompt already gets you there. Costs are real: data prep, training compute, eval, and re-training when the base model upgrades.

Example

Acme fine-tunes a small model on 5,000 of their best support replies. The fine-tuned model writes in their voice cheaper and faster than prompting the base model.

How OpenLabor uses it

OpenLabor uses fine-tuning sparingly — usually only for high-volume voice-matching tasks where prompting alone falls short.

When should I fine-tune?

When the prompt is already great and you want to capture stable style at lower cost. Almost never as the first move.

Related: llm, rag, embedding.

AI Labor Glossary