principles.fyi · the brain · concept
fine-tuning
Extra training that teaches a pretrained model to actually answer.
Pretraining (predict the next word on the whole internet) leaves a model that knows a lot but just continues text — it does not answer questions. Fine-tuning continues training on curated example conversations (a question paired with a good answer), so the model learns the assistant format: when asked something, reply helpfully. LoRA is one cheap way to do this.
Appears in
- How it learns Transformers, ELI5 · pt 8
- Put it to work Masked Language Models · pt 5
- Talking to it — prompting LLMs in the Wild · pt 2