Models

Two tracks. No trained model yet.

Each model gets a card here once it exists. Until then this page shows the plan and what has been measured on the candidates.

Track B · fine-tune

A reviewer for apps built with AI tools

A LoRA adapter on a small open-weight model. The task is reviewing and repairing apps built on Supabase: missing or wrong row-level security policies, missing ownership checks, exposed keys. It has to work in English and in Arabic.

not trained yet

Table 2 · parent candidates

Qwen3.5-4BGemma 4 E4B
LicenceApache-2.0Apache-2.0
How it was loaded16-bit4-bit, with the embedding table and some layers in system RAM
Peak VRAM, first load8.72 GiB7.09 GiB
Peak VRAM, LoRA at sequence 5128.81 GiB7.09 GiB
Training time, 100 steps159 s231 s
Levantine Arabic, one promptanswered in EgyptianLevantine, judged natural
Vulnerable trials repaired, untuned18 of 42 (43%)not measured yet
Clean apps left working, untuned8 of 9 trialsnot measured yet

The first load used four prompts per model, which is a sanity check and does not decide the parent. The repair eval is a pilot: 17 tasks, 3 trials each. A third candidate, Jais 3, has not been tried.

What counts as a result

  • An adapter has to beat the untuned parent’s 43% by about 30 points at this task count. Anything smaller is reported as no measurable difference.
  • It also has to beat a stronger model with lint tools on the same tasks. That comparison has not been run.
  • It must not break the clean apps.

Track A · from scratch

A 17M-parameter model, built whole

A decoder-only Transformer trained from random weights with our own tokenizer. The goal is to learn the full stack: tokenizer, pretraining, checkpointing, sampling. It will not be a usable product model.

not trained yet

Layers8 blocks, width 384, 6 heads
Parameters17,301,504, counted by hand
Context512 tokens
TokenizerByte-level BPE. Vocabulary size is still open.
Tokens per second on this cardnot measured yet