Choose a training method

Training supports two methods: supervised fine-tuning (SFT) and prompt gisting. Both learn from conversations, but they solve different problems and produce different artifacts.

Compare the methods

Question	Supervised fine-tuning	Prompt gisting
What changes?	The model’s behavior is adapted from labeled conversations.	A static prompt is replaced by learned soft tokens.
What input is required?	A frozen dataset slice of conversations with a split column.	The same dataset slice and a non-empty prompt text file.
What does it produce?	A LoRA adapter that runs with its Base model.	A full model containing the learned gist tokens.
Which catalog Base models are supported?	`qwen3-4b`, `qwen3-8b`, `qwen3-32b`, and `qwen3-235b`.	`qwen3-4b` and `qwen3-8b`. The larger full models exceed Model Version size limits.
Can it continue from an existing Model Version?	Yes, from a LoRA adapter Model Version in the same project.	Yes, from a LoRA adapter Model Version in the same project.
Main method-specific setting	LoRA rank: `8`, `16`, or `32`.	Gist tokens: `8`, `32`, `64`, or a supported fraction of the prompt.

Both methods publish a downloadable Model Version. The platform does not host or serve that Model Version.

Choose SFT

Choose supervised fine-tuning when examples should teach a model how to respond or classify. Each row ends with the assistant output the model should learn, such as a support-ticket label or a complete answer. SFT is also the method to use when:

You need a LoRA adapter rather than a full model.
You need one of the larger qwen3-32b or qwen3-235b Base models.
You want to continue adapting an earlier LoRA adapter.

SFT is the CLI default. You do not pass --method when creating an SFT Training.

Choose gisting

Choose gisting when the reusable part of the task is a long, static instruction prompt. Gisting learns soft tokens from that prompt and the example conversations. The dynamic user input stays in each conversation. Gisting is also the method to use when:

Prompt compression is the goal, rather than adapting model weights for a new response behavior.
You can use qwen3-4b or qwen3-8b.
You need a full Model Version rather than a LoRA adapter.

A gisting Training always needs --method gisting and --prompt-file.

Use the right product surface

Use the CLI or HTTP API for project Training. The web app’s Fine-tuning and Prompt gisting pages help you set up a coding agent. Fine-tune lab, where available, has separate runs and downloads; see Ways to work with Training.

Next steps

Train your first model for a guided SFT Training Run
Fine-tune a model for the complete SFT CLI workflow
Compress a prompt with gisting for the complete gisting CLI workflow
Training settings for every configurable value and default
Base models for the model catalog and pinned revisions

Start

Data Labeling

Train

Workflows

Admin

Compare the methods

Choose SFT

Choose gisting

Use the right product surface

Next steps

​Compare the methods

​Choose SFT

​Choose gisting

​Use the right product surface

​Next steps

Compare the methods

Choose SFT

Choose gisting

Use the right product surface

Next steps