rchaz/tunelab
6 stars · Last commit 2026-07-25
Claude Code plugin for LLM fine-tuning, distillation, and evaluation — decide whether you need fine-tuning at all, distill your LLM logs into small local models (MLX/LoRA), evaluate with held-out discipline, and learn the why at every step.
README preview
<p align="center">
<img src="assets/tunelab-banner.png" alt="tunelab — move repetitive LLM calls (tool calling, classification, extraction) onto small local models you fine-tune on your own data with MLX/LoRA, verified on held-out data before you ship" width="100%">
</p>
# Tunelab
tunelab moves repetitive LLM calls — tool calling, classification, extraction — onto small local models. It preps and labels training data from your logs (or distills it from a bigger model), fine-tunes locally with MLX/LoRA, and verifies the small model beats the API on held-out data before you ship.
Measured on Banking77: a free local classifier scores **88.5%** vs Claude Opus 4.8's **81.8%**; a 3-tier cascade reaches **94%** at **8× lower cost** than frontier-only.
```
You: We classify support tickets with Claude Opus 4.8. It costs $40/day. Can we do this cheaper?
tunelab: Trained a free local classifier on your 8k tickets.
Free model: 88.5% — Claude Opus 4.8: 81.8%
The free one is more accurate. No fine-tuning needed.
```
```
You: The model outputs random text. I need it to return exact label names like "card_arrival".