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.*