About this research

An independent project on what small models - one to a hundred million parameters - can do when they own one narrow task instead of losing to a frontier model at everything. Everything on this site is an instrument over a model that was trained here, from scratch, and every study page ends with what the result does not show.

from-scratch GPT · PyTorchmulti-seed comparisons, CUDA-measurednulls published, not deleted

How the numbers are made

The house rules, because they are what make the rest of the site worth reading:

The four strands

Who did this

iksnerd, independent researcher, no affiliation. Everything on this site - the models, the instruments, the measurements and the write-ups - is one person's work, which is why the house rules above are written down rather than assumed.

What is not on this site

The full write-ups. Each strand above has a findings document behind it - the tables, the seed spreads, the caveats - and the source repository is private for now. The model weights are being published on Hugging Face under glassbox - the roster page lists what's live and what's still on its way. The architecture log is published in full at /findings.

What is already downloadable: the Bulgarian small-model toolkit - the 91M flagship, the 14M model as int8 ONNX, the restorer with both its WASM builds, the pair set, and the browser demo. The numbers those repos are judged on are re-runnable rather than quoted: bg-eval is a standalone MIT harness that loads the published checkpoints without this project's training code, and its expected values are checked into the repo, so a mismatch is visible rather than arguable.

The mark

Glassbox's icon, in the six appearances iOS renders an app icon through: flat, dark, tinted, and behind glass.

Glassbox app icon, default appearance
Default
Glassbox app icon, dark appearance
Dark
Glassbox app icon, tinted appearance
Tinted
Glassbox app icon, tinted, dark appearance
Tinted, dark
Glassbox app icon, clear appearance
Clear
Glassbox app icon, clear, dark appearance
Clear, dark
Built and measured by iksnerd. The instruments here are the same ones used to do the work, not a presentation layer built afterwards.