Fine-tuning is training a pretrained model further on your own data so that it behaves the way a task needs. It changes the model's weights, which is what sets it apart from prompting.
How it works
A pretrained model already knows language. Fine-tuning runs more training steps on a much smaller dataset: the model predicts, a loss measures the error against your examples, and gradients update the weights. The run is short next to pre-training, so the model keeps most of what it knows and shifts toward the new data.
What gets updated is a choice:
- Full fine-tuning updates every weight.
- LoRA (low-rank adaptation) and other parameter-efficient methods (PEFT) freeze the model and train a small set of added weights.
What it is for
- A fixed output format, tone or style.
- A narrow task where a small model has to match a large one.
- A domain with its own vocabulary, such as law or medicine.
- Behavior that a prompt cannot hold reliably, or a prompt too long to send on every call.
Fine-tuning does not keep facts current. Retrieval-augmented generation (RAG) does: see fine-tuning vs RAG.
In Halo
Halo is our open-source framework for post-training open models, from supervised fine-tuning to reinforcement learning. One command starts a LoRA run on one GPU:
Bash
halo launch sft examples/sft/qwen3/qwen3-4b-ultrachat-lora.yamlThe quickstart goes from a running container to a finished run.
