jevlike

Jevlike is an independent, open implementation of the input and output shape a Jev-style model uses: a piece of text plus a changing list of options in, one probability per option out, computed in a single forward pass instead of by generating text token by token. The README notes that TypeSafe has not published Jev’s design, so this is an independent starter with the same input and output shape, not a copy of Jev.

The scorer treats each option as a query vector that attends over the context tokens, producing one context vector per option. A shared dot product turns each option and context pair into a score, and a softmax across the options turns the scores into probabilities. The default encoder, the part that turns text into vectors, learns byte embeddings from scratch; you can also swap in a frozen pretrained Hugging Face encoder and train only a small scorer head on top of it.

The repo includes a Doom demo and a chess demo that reuse the same option-attention head to score controller buttons from image patches instead of text, treating each button as an option. The author is explicit that the demo clips were chosen for activity, not as claims about typical play or skill.

Training and evaluation run on JSONL files of context, options and a correct-answer index. The eval command reports top-1 and top-3 accuracy plus expected calibration error, measured against a shuffled-context control so a model can be checked against a baseline that never sees the real pairing.

Open on GitHub