File size: 1,642 Bytes
be91b8c 60709b9 be91b8c 60709b9 8ed4343 60709b9 8ed4343 60709b9 d07c1e8 37a1dd1 d07c1e8 8ed4343 d07c1e8 8ed4343 d07c1e8 8ed4343 d07c1e8 8ed4343 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 | ---
title: README
emoji: 🧪
colorFrom: blue
colorTo: indigo
sdk: static
pinned: false
---
# MAIN Lab · Master Artificial Intelligence Network
## Big AI trains small AI.
**Distilling frontier intelligence into small models.**
Frontier models are the teachers. Small models are the students: fast, cheap, private, and able to run offline on a laptop or a phone.
At MAIN Lab, frontier AI agents (**Claude Code** and **Codex**) do the hands-on work: they build the datasets, run the training, edit the weights and run the benchmarks. A human directs the research and judges the results.
We start with **Azerbaijani**, a language big models still underserve.
### Open by default
- **Open weights:** every model we train is released here.
- **Open results:** the same tests for every model. Wins, losses and ties all get published.
- **Small models:** everything runs on your own hardware.
### Models
| Model | Method | What changed |
|---|---|---|
| [gemma-3-270m-az](https://huggingface.co/MAIN-Lab/gemma-3-270m-az) | Fine-tuning, CPU only | Azerbaijani grammar 51.5 → 80, EN→AZ translation chrF++ 14.0 → 25.1 |
| [qwen2.5-0.5b-az-steer](https://huggingface.co/MAIN-Lab/qwen2.5-0.5b-az-steer) | Weight-space programming, no training | 128 numbers changed: Turkish answers 42% → 18% |
### Benchmarks
1. **#001:** Small models, Azerbaijani (a 5-model leaderboard)
2. **#002:** Weight-space programming for math (a negative result: a tie)
3. **#003:** Weight-space programming vs Turkish drift (it worked)
### Try it
- [Before / after answers](https://huggingface.co/spaces/MAIN-Lab/az-steer)
*Small models. Frontier teachers.*
|