NAME0x0 · Dubai
Systems & ML engineer. I build systems under real constraints.
Ideas tend to arrive together. I get caught up in finding the connections, following them across fields, and seeing what I can bring into the same project.
Now AVA v3, a coding specialist, in research
Contact Open to systems and ML work · Dubai or remote
Same particles. Different states. Read
Work
Windows desktops and shells, native tools, a language, web experiments, trading agents and small models.
All 18 projects
In depth AVA · How much can a small model learn?
Models of 11M to 99M parameters, trained on a laptop GPU with 4 GB of VRAM. The bare 14M baseline scored about 24% on ARC-Challenge. Guessing scores 25%.
Retrieval, memory and tool scaffolds joined the model. A retrieval ensemble reached 91 of 299. The limit was knowledge: a model this small cannot store enough of it.
Qwen3.5-2B does not fit the card at 16-bit. Loaded in 4-bit NF4, it does. Its weights stay frozen; only a thin QLoRA adapter trains.
AVA v1 trained on 5K examples for one epoch: 66% on ARC-Challenge, exactly the base model. A completed run was not an improvement.
A 20K curated corpus of maths, science, reasoning and tool examples. Peak training VRAM: 1.81 GB. ARC-Challenge: 82.0% on all 1,172 questions.
AVA v2 ships as a 42 MB adapter with quantized GGUF builds. The Gemma 4 branch is paused; v3 is in research. Maths, narrative commonsense and tool use remain weaker.
Repository-reported. Caveats
AVA began with small models trained from scratch. Tokenisers, retrieval and memory all became part of the experiment. The models reached a knowledge ceiling; the useful parts stayed.
Keep the question. Reconsider the starting point.
The next approach adapted a pretrained model with QLoRA. It made training on the laptop practical, but the first fine-tuning attempt did not improve ARC reasoning over the base model.
A working training run was only the beginning.
A larger, curated mixture of science, maths, reasoning and tool examples followed. Training optimisations and a broader evaluation pipeline grew alongside it.
Improve the material. Check the result.
AVA v2 became a downloadable adapter and local model. The experiment history keeps the failed and paused branches too: a record of what was tried, what shipped, and what is still open.
A release is a place to continue from.
Person
Ideas tend to arrive together. I get caught up in finding the connections, following them across fields, and seeing what I can bring into the same project.
Some moments from outside the repositories.
Questions
Separate repositories, recurring questions. The body takes each project's shape in turn; hover a question to choose one. These projects do not depend on one another, but they keep returning to limits, choices and coordination.
Explore
Six rooms drawn in characters, with a rolling point cloud at the centre. Follow the gates, read the exhibits, collect the torn fragments. Exit at any time; the world starts fresh each visit.
Arrow keys or W A S D · touch controls · map · Escape to leave
