Install

Needs Node.js 18+ and a running Ollama with at least one model pulled.

From npm

npm install -g finetuna
finetuna --help

Or:

pnpm add -g finetuna

Look first, then create

First inspect a model you already have. Create the named variant with the shown settings. Then verify that variant.

finetuna --check
finetuna --check --model llama3.2 --json
finetuna --model llama3.2 --name llama3.2-ft --ctx 32768 --auto-tune
finetuna --verify llama3.2-ft

llama3.2 is an existing model you substitute. llama3.2-ft is the name from --name in the create step. --check and --dry-run never run ollama create. --verify <name> loads that existing name and reads /api/ps. It does not create a new model. Re-run it after you change the model, context, concurrency, or host load. A pass is a snapshot of that loaded run, not a perpetual residency guarantee.

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/

FINETUNA_DIR overrides the data directory. OLLAMA_HOST points at another Ollama.

Auto-tune

Default is context fit-search. --tune-batch opts into a num_batch sweep. Selection is median + spread: a difference inside the noise keeps the incumbent.

Client presets: --openclaw / --hermes / --continue. Flash attention is an Ollama server setting (OLLAMA_FLASH_ATTENTION=1), not a Modelfile parameter.

Full flags, env vars, and Apple Silicon notes: README on GitHub.