Generate an Ollama Modelfile on export

[will-lamerton]

Problem

After nanotune export you have a GGUF file and no guidance on running it anywhere other than Nanotune. Ollama is the most common destination for a local GGUF, and getting a fine-tune into it requires hand-writing a Modelfile — including re-deriving the system prompt and template, which Nanotune already knows from config.json.

Proposal

nanotune export --ollama writes a Modelfile next to the GGUF:

FROM ./my-model-q4_k_m.gguf
SYSTEM """<contextMessage.content from config.json>"""

Print the follow-up command on success:

ollama create my-model -f .nanotune/models/Modelfile

Acceptance criteria

• Modelfile is generated with the project's system prompt embedded
• Path in FROM is correct relative to the Modelfile
• Success output tells the user exactly what to run next
• Documented in docs/commands/export.md

Notes

Good first issue — one template string, one file write, one docs update. Everything it needs is already loaded in src/commands/export.tsx via loadConfig() and resolveContextMessage().

[yashksaini-coder]

Would like to take this one. Approach: a pure buildModelfile() in src/lib/ that export --ollama writes beside the GGUF, so the generation is unit-testable on its own rather than behind the Apple-Silicon-gated export path.

Another thing is that the Modelfile format has no escape mechanism — unquote() in Ollama's parser/parser.go returns whatever sits between the delimiters verbatim — and a """ block terminates at the first line ending in """, so a system prompt containing that (plausible for a Python fine-tune) silently truncates and turns the rest into invalid commands. I plan to mirror Ollama's own quote() rule exactly, and fail with a clear message on the one pattern that can't be represented, rather than write a file that errors confusingly at ollama create time.