Ask HN: Do You Need a CLI Tool for Output Schema Fine-Tuning of Open Source LLM
Who's deploying Open Source models and would like a simpler way to fine-tune the models out of HuggingFace / Ollama?
I was going to build a tool for me because I have a lot of Agent fine-tuning (and re-tuning) required. If there's interest, I can share my code / learning.
The minimum most of my agents from the 'base' models is training them on output format schema so they're more consistent.
I want the process to be simple enough that I could put them reliability in a build & deployment pipeline.
Vision of how it'll work:
(setup) 0. `pip install [newlibary]` & run a setup command on that library
(day-to-day usage) 1. '[newlib] create training_manifest.yaml' (or json, feedback welcome)
2. '[newlib] tune `model-name`' (name in manifest)
3. '[newlib] verify `model-name`'
When we verify we should be able to see the improvement in results on a set of verification tests.
I'll be using Ollama behind the scenes, we should, eventually be able to push and pull our new models to a store.