Install
Requires Node.js 18+ and a running Ollama with at least one model pulled.
From npm
npm install -g finetuna
finetuna --help
Or with pnpm:
pnpm add -g finetuna
Look first, then create
finetuna --check # will it fit? how much context?
finetuna --check --model llama3.2 --json
finetuna --verify my-model-ctx32k
finetuna # interactive create
finetuna --model llama3.2 --name llama3.2-ft --ctx 32768 --auto-tune
--check / --dry-run never run ollama create.
From GitHub
npm install -g github:Catalyst-Forge-LLC/finetuna
Where files go
| Mode | Modelfile | State / results |
|---|---|---|
Checkout (pnpm start) | ./Modelfile-finetuna | current directory |
Installed (finetuna on PATH) | ./Modelfile-finetuna (cwd) | ~/.finetuna/ |
Override the data directory with FINETUNA_DIR. Point Ollama at another host with OLLAMA_HOST.
Auto-tune
Default auto-tune is context fit-search. Phase 1 (num_batch) is opt-in via --tune-batch. Selection uses median + spread — differences inside the noise keep the incumbent.
Client presets: --openclaw / --hermes / --continue. Flash attention is an Ollama server setting (OLLAMA_FLASH_ATTENTION=1), not a Modelfile parameter.
Full flag table, env vars, and Apple Silicon notes: README on GitHub.