SAURON — a LoRA adapter trained by its own fleet
SAURON is an experimental LoRA adapter for Qwen3-4B (4-bit), trained by the TibiOS fleet itself: five ordinary home machines that generated their own training data, trained the adapter, and deployed it back as a new inference lane. Zero cloud, zero token spend.
The fleet that trained it
| Machine | RAM | Role in training |
|---|---|---|
| Mac mini M4 | 16 GB | Training (MLX, Apple Silicon) |
| MacBook Air M1 | 16 GB | Training (MLX, Apple Silicon) |
| 2 × NVIDIA Jetson Orin Nano | 8 GB each | Data generation + training (Unsloth + TRL, CUDA) |
| Orange Pi 6 Plus | 16 GB | Head node / router — publishes the adapter |
All of it over a home Wi-Fi LAN.
The loop
- The fleet's inference lanes generate their own training data (Q&A glossary pairs about TibiOS).
- The machines train in a ring, splitting each batch by every machine's measured speed.
- Training produces this LoRA adapter.
- The head node publishes it as a new inference lane (
@tuned). - The fleet infers again — and can generate data for the next cycle.
Adapter specifications (this file)
- Base model: Qwen3-4B, 4-bit quantized
- Type: LoRA, rank 8, scale 20 (MLX format)
- Coverage: last 8 transformer layers, 7 projections each (q, k, v, o, gate, up, down)
- Trainable parameters: 3.67M (~0.09% of the base model)
- File size: 14 MB, float32, safetensors
- Training: 1,294 iterations, then resumed +600 with a presentation Q&A block (English and Spanish) so the model can answer "What is TibiOS?"; lr 1e-4 → 5e-5, sequence length 512, batch 4, seed 7
Usage (MLX)
pip install mlx-lm
mlx_lm.generate \
--model mlx-community/Qwen3-4B-4bit \
--adapter-path . \
--prompt "What is TibiOS?"
Example output:
TibiOS is a network of home computers pooling their compute to train and serve AI models, without the cloud.
It answers glossary-style questions best (the training data is a Spanish Q&A glossary of the TibiOS architecture). For example, "¿Qué representa la Runtime API en el sistema de TibiOS?" gets:
Runtime API es la única superficie pública y coherente mediante la cual un consumidor externo —una aplicación, el SDK, el CLI u otro Runtime de TibiOS— puede interactuar con la Runtime. No es un protocolo ni un transporte, solo convierte en accesibles desde el exterior las operaciones que ya existen dentro de la Runtime.
Honesty notes
- This is an experiment, not a production model: it was trained on a small glossary-style dataset and its facts are often wrong. The point is the loop — the same fleet trains and serves — not the adapter's knowledge.
- Kimi K3 and Claude were used exclusively to draft the glossary data, not to train SAURON.
- Training infrastructure: TibiOS · Landing + demo video
Contact
fgparamio@tibios.ai · +52 999 122 3748
Quantized