README / README.md
Turan0k's picture
We train models, not build them
37a1dd1 verified
|
Raw History Blame Contribute Delete
1.64 kB
metadata
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 Fine-tuning, CPU only Azerbaijani grammar 51.5 → 80, EN→AZ translation chrF++ 14.0 → 25.1
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

Small models. Frontier teachers.