Teach a model
your style.
A small local model learns from your rated answers, notes and files. An adapter carries what it learned into your chats.
What training does
Open Customize → Models → Training. Four guided steps train a small model on this computer and turn the result into an adapter: a small file that sits on top of the model, so the model itself is unchanged.
Training teaches how answers should sound and look. For facts, a project's knowledge files work better. See Projects.
Training is new, and so far only tested with stand-ins for the GPU parts. Expect rough edges, and check the result with the comparison in step four before you rely on it.
What you need
- An NVIDIA GPU with 6 GB or more, for models up to about 3B. This computer, at the top of the page, shows what training can use here and how much disk space is free.
- The training tools. The first time, press Install the training tools. You can remove them later; trained adapters and datasets stay.
Without a usable NVIDIA GPU, training would run on the processor and take hours. If you have one, check that its driver works by running nvidia-smi in a terminal, then come back. On a Mac, the guided steps can't load the result into llama.cpp yet; the advanced form trains with MLX.
The four steps
- Data. Pick what the model learns from: chat answers you marked good with the thumbs up, each with the conversation before it; your notes, as plain text; and text or Markdown files in a project's folders. Give it a name and press Build training data. With fewer than 30 examples, expect a small change. A few hundred teach much more.
- Model. See which models fit your GPU (Fits, Tight fit or Too big for this GPU), roughly how long each takes, and what it downloads first. The one that suits your card is marked Suggested. Times are rough guesses, and the first run also spends a few minutes loading.
- Train. Press Start training, or download and train in one go. The step shows progress as it goes, and Stop ends it.
- Use it. Press Make the adapter, then load the model with it into llama.cpp. Under Ask both, type a question and press Compare to see the answer without the adapter and with it, side by side.
The models offered here are the same ones the first-run setup offers to run on this computer, so an adapter you train fits a model you may already chat with.
Your adapters
Trained adapters are listed under Your adapters. A switch on each decides whether it loads with its model. Turned off, it is left out from the next time the model loads. You can delete an adapter when it isn't loaded; the training job and its data stay.
Datasets, adapters and what training runs left behind are also in Files → Training, with their sizes. See Files.
Advanced training
Below the guided steps, Advanced: train any model on any dataset, with every setting takes a model and a JSONL dataset, with settings such as epochs, learning rate, LoRA rank and alpha, batch size and maximum sequence length. Each job shows its metrics and logs.
Testing training on more GPUs, and on Macs, is in active development. See Project direction.