GitHub · Triage and routing · Research and data
TianyuCodings/NanoJev
The smallest fully-open reference implementation of the architecture.
What it does
A 0.6B "nano replica" of the Jev idea: states and questions in, complete probability distributions out, with zero output-token decoding. It ships the training pipeline, the results, a Hugging Face model and dataset, and playable browser demos including ViZDoom, maze and snake.
The repository's own description: “A nano replica of Jev: parallel decisions, dynamic candidates, and an end-to-end training pipeline.”
How it works
Parallel decisions over dynamic candidate sets; `configs/` and `scripts/` drive training, `results/` holds the runs, `web/` the demos, and `SOURCE_MANIFEST.json` records what the authors derived from elsewhere — the kind of file most reimplementations do not publish.
The repository, by the numbers
Read from the GitHub API on 2026-09-22. Stars and the last push move daily — quote them with the date, the way we do.
What we checked
- the tree: configs, results, research, scripts, web, SOURCE_MANIFEST.json
- MIT LICENSE (OpenJev contributors)
- README states 0.6B, zero output-token decoding, and links the model and dataset on Hugging Face
- the README links a ViZDoom demo that runs with autoplay
What we did not check. We did not run it; the numbers in results/ are the authors' and we did not re-derive them.