scatha¶
Train and fine-tune open language models in Go.
scatha does full training and parameter-efficient fine-tuning (LoRA, QLoRA) of open models such as the Llama, Qwen and Mistral families, on XLA through GoMLX without cgo. It reads Hugging Face checkpoints and writes full checkpoints and PEFT adapters that the Python ecosystem, vLLM included, loads as its own. Its companion CLI, dunscaith, submits and supervises the jobs that run it.
Design phase
scatha exports nothing yet. Decisions are being written as specs in
the project wiki, and nothing is
implemented before its spec is APPROVED.
The name¶
Scáthach (say "SKAH-hakh") ran a school of arms on the Isle of Skye in the Ulster Cycle. Heroes who were already warriors, Cú Chulainn and Fer Díad among them, came to her to be finished: taught feats nobody else knew. Specialising an already-trained model in a narrow skill is fine-tuning; scatha also does the training that comes before it.
Documentation¶
The documentation follows the Diátaxis framework:
- Tutorials: learning-oriented lessons.
- How-to guides: task-oriented recipes for specific goals.
- Explanation: how scatha works, and why.
- Reference: the API, configuration and file formats.