Fernando Gutierrez
SAURON gets its own page: the real self-training story; home carries a teaser
5322285 Download sauron.html from fgparamio/TibiOS: direct link, hf CLI and curl.
- Browser
- Download file 4.67 kB
-
https://huggingface.co/spaces/fgparamio/TibiOS/resolve/main/sauron.html
- Command line
-
hf download hf://spaces/fgparamio/TibiOS/sauron.html
-
curl -L -o sauron.html https://huggingface.co/spaces/fgparamio/TibiOS/resolve/main/sauron.html
4.67 kB
| <html lang="en"> | |
| <head> | |
| <meta charset="utf-8" /> | |
| <meta name="viewport" content="width=device-width, initial-scale=1" /> | |
| <title>SAURON — a model trained by its own fleet · TibiOS</title> | |
| <meta name="description" content="SAURON: a Qwen3-4B LoRA adapter trained by the TibiOS fleet itself — five home machines, zero cloud, zero token spend." /> | |
| <link rel="icon" href="assets/tibios-logo.png" /> | |
| <link rel="stylesheet" href="style.css" /> | |
| </head> | |
| <body> | |
| <header class="nav"> | |
| <div class="brand"> | |
| <img src="assets/tibios-logo.png" alt="TibiOS logo" class="logo" /> | |
| <span>TibiOS</span> | |
| </div> | |
| <nav> | |
| <a href="index.html#how">How it works</a> | |
| <a href="index.html#fleet">The fleet</a> | |
| <a href="index.html#demo">Demo</a> | |
| <a href="sauron.html">Sauron</a> | |
| <a href="index.html#contact">Contact</a> | |
| </nav> | |
| <a class="btn btn-small" href="https://tibios.ai">tibios.ai</a> | |
| </header> | |
| <section class="hero"> | |
| <p class="tagline">Presenting</p> | |
| <h1>SAURON</h1> | |
| <p class="tagline">A model trained by its own fleet.</p> | |
| <p class="lede"> | |
| What if a set of home computers could generate their own training | |
| data, train a model, deploy it and use it for inference — all within | |
| the same infrastructure? That is what we achieved with TibiOS. | |
| <strong>Zero token spend.</strong> | |
| </p> | |
| </section> | |
| <section class="section"> | |
| <h2>The fleet that trained it</h2> | |
| <p class="lede center"> | |
| SAURON is based on Qwen3-4B, adapted with LoRA on a heterogeneous | |
| fleet of five machines: | |
| </p> | |
| <div class="cards"> | |
| <div class="card"><h3>🍎 Mac mini M4</h3><p>16 GB — training with MLX.</p></div> | |
| <div class="card"><h3>💻 MacBook Air M1</h3><p>16 GB — training with MLX.</p></div> | |
| <div class="card"><h3>⚡ 2 × Jetson Orin Nano</h3><p>8 GB each — training with Unsloth + TRL on CUDA.</p></div> | |
| <div class="card"><h3>🔌 Orange Pi 6+</h3><p>16 GB — head node and router.</p></div> | |
| </div> | |
| </section> | |
| <section class="section"> | |
| <h2>The full pipeline</h2> | |
| <div class="steps"> | |
| <div class="step"><span>1</span><p>The inference lanes generate their own training data.</p></div> | |
| <div class="step"><span>2</span><p>The Macs train in a ring, splitting each batch by every machine's measured speed.</p></div> | |
| <div class="step"><span>3</span><p>Training produces the SAURON LoRA adapter.</p></div> | |
| <div class="step"><span>4</span><p>The head node publishes the adapter as a new inference lane: <strong>@tuned</strong>.</p></div> | |
| <div class="step"><span>5</span><p>The fleet infers again — and can generate new data for the next cycle.</p></div> | |
| </div> | |
| </section> | |
| <section class="section"> | |
| <h2>Technical data</h2> | |
| <div class="cards"> | |
| <div class="card"><h3>Qwen3-4B · 4-bit</h3><p>Base model, quantized.</p></div> | |
| <div class="card"><h3>MLX + Unsloth/TRL</h3><p>MLX on Apple Silicon; Unsloth + TRL on CUDA.</p></div> | |
| <div class="card"><h3>LoRA r=8, α=16</h3><p>Rank 8, alpha 16, seven projections.</p></div> | |
| <div class="card"><h3>16.5M parameters</h3><p>Trainable — about 0.74% of the model.</p></div> | |
| <div class="card"><h3>64 MB</h3><p>Adapter size.</p></div> | |
| <div class="card"><h3>Home Wi-Fi LAN</h3><p>Zero electricity cost in our scenario, no cloud needed for this local pipeline.</p></div> | |
| </div> | |
| <p class="lede center note"> | |
| Kimi K3 and Claude were used exclusively to draft the glossary data — | |
| not to train SAURON. | |
| </p> | |
| </section> | |
| <section class="section"> | |
| <img class="wide" src="assets/sauron-fleet-training.png" alt="SAURON infographic: the fleet, the pipeline and the LoRA adapter data" /> | |
| </section> | |
| <section class="section"> | |
| <p class="lede center big-quote"> | |
| The important result is not just a tuned model. It is a working cycle | |
| in which the same fleet both trains and infers — a fundamental piece | |
| of the TibiOS vision: distributed AI infrastructure where heterogeneous | |
| machines collaborate, learn and serve models. | |
| </p> | |
| <p class="lede center big-quote"> | |
| <strong>You don't need to start with a data center.<br />Start with what you have.</strong> | |
| </p> | |
| </section> | |
| <footer id="contact" class="footer"> | |
| <img src="assets/tibios-logo.png" alt="" class="logo" /> | |
| <p><strong>TibiOS</strong> — Decentralized AI for everyone</p> | |
| <p> | |
| <a href="https://tibios.ai">https://tibios.ai</a> · | |
| <a href="mailto:fgparamio@tibios.ai">fgparamio@tibios.ai</a> · | |
| +52 999 122 3748 | |
| </p> | |
| <p class="tiny">Global community — we collaborate, we grow, we innovate.</p> | |
| </footer> | |
| </body> | |
| </html> | |