Tinker finetuning

[johny-b]

Add tinker finetuning.
Ideally, keep the same finetuning interface - if possible.

[johny-b]

Looked into this, looks a bit hard :/

OpenAI finetuning is 'fire and forget', tinker finetuning requires a running client script I think. So these two are totally different and keeping the same interface might be hard.

So probably it makes no sense to try keeping similar interface. Instead, just have a tinker finrtuning script that:

• Uses the same (or similar) files.jsonl, for files utilities
• Maybe uses jobs.jsonl file if useful
• Adds the final sampler to models.csv?

maybe that's good enough?

[johny-b]

Idea: make default batch size such that we have e.g. 1000 steps.