Tokenizer Fine-Tuning Repo (darcysmithcollabs)

1. generate_data.py — Synthetic Q&A pairs via OpenAI-compatible endpoint 2. filter_data.py — Dedup + quality filter → data/train.jsonl 3. train.py — LoRA fine-tune (peft + trl) → adapters 4. evaluate.py — Base vs tuned comparison → cache/comparisons.json 5. merge_and_interp.py — Merge adapter + TransformerLens interpretability → cache/interp/ 6. notebook.py — The marimo presentation

A "Noosa Council service assistant": synthetic resident enquiries (rates, bins, permits, noise, beaches, trails) generated by a local LLM, filtered, and used to LoRA-fine-tune Qwen/Qwen2.5-3B-Instruct into a consistent plain-language council voice. Fully local — no data leaves the building.

Darcy is using this repo as a teaching onramp for their swimming-tech AI collaboration. The fine-tuning pattern (generate domain-specific data → filter → LoRA → evaluate → interpret) is directly applicable to training a swimming analysis model.