Should we publish public fused GGUF / Ollama weights?

[YauhenBichel]

Context

The GitHub repo is the source of truth for code + harness. Weights are meant to live on Hugging Face: YauhenBichel/python-vibe-0.5b.

A 400-iter MLX QLoRA run overfit after ~step 100 (val ~0.91 then ~1.63 at 400). Shipping a fused 4-bit blob of the last checkpoint would freeze a worse model than the mid-run adapter.

Questions

1. Publish adapters only until val loss is stable?
2. Publish a fused GGUF / Modelfile so ollama pull works without MLX?
3. Keep the Hub repo private until a held-out eval exists?

Constraints we will not drop: no real keys in the card, no curl|sh examples, Apache-2.0 on the code.

Comment with what you would actually download and run.