jevlike
Training library for a small one-pass model that scores a changing list of text options and returns a probability per option.
Overview
Jevlike lets you train a small model with the same input and output shape as Jev: a piece of text plus N text options in, one probability per option out, in a single forward pass rather than generated text. It is an independent starter model, explicitly not a reproduction of TypeSafe's unpublished architecture.
Each option becomes a query that attends over the context tokens to build an option-specific context vector, a shared dot product scores each pair, and a softmax runs across options. The default encoder learns byte embeddings from scratch, with an optional frozen Hugging Face encoder. The same option-attention head is used in visual demos that score Doom controller buttons and chess moves from image patches, and a Wikispeedia next-click example. The README is candid about demo strength: the chess checkpoint drew or beat a random mover but lost 48 of 50 to Stockfish level 0. Other projects, such as CUA-S1-FORMS and an ESP32 port, build on it.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which of these N text options best fits the context? | — |
| Choice | Which controller button should be pressed in this game frame? | — |
| Choice | Which link should be clicked next to reach the target article (Wikispeedia)? | — |
How it works
- Prepare examples of context, option list and correct option
- Train the encoder and option-attention head
- At inference, score each option against the context in one pass
- Pick the argmax or threshold on the softmax probabilities
Strengths
- Small, trainable and self-hosted
- Handles variable-length option lists
- Same head works for text and image-patch inputs
- Honest reporting of demo limits
Caveats
- Starter model, not a production decision model
- Game demos are weak against real opponents
- Quality depends on your own training data
Getting started
uv pip install -e '.[games]' then run the Doom or chess examples under examples/.At a glance
- Type
- Library
- Status
- research
- Language
- Python
- Runtime
- PyTorch
- Has code
- yes
Use cases and integrations
How curated lists describe it
Training library: build a small model that chooses among a changing list of text options and returns one probability per option in a single pass - the base CUA-S1-FORMS was built on.
yibie/awesome-jevTrain a small one-pass scorer that maps context + N text options to a probability per option. Includes Doom / chess vision demos and a Wikispeedia next-click example. Explicitly *not* a reproduction of TypeSafe's architecture or RLCD.
AnotiaWang/awesome-decision-modelsTrainable encoder and option-attention head for variable candidate sets, with separate visual game experiments.
cobanov/awesome-jevTrain a small model that chooses among a changing list of text options, one probability per option in a single pass. Includes Doom, chess, and Wikispeedia demos.
hellogumbo/awesome-jevTrain a small model that scores a changing list of text options in one pass.
MrJev/awesome-jev
Frequently asked questions
Is jevlike a copy of Jev?
No. It shares Jev's input/output shape but uses its own option-attention architecture.
What can I train it to choose?
Any task framed as picking among a changing list of text options, such as routes, links or game actions.
Does it need an API key?
No, you train and run it locally.
Related projects
Python library that turns an open LLM into a Jev-style classifier via constrained single-token choice indexes in one forward pass.
ProjectprototypePythonOpen family of System One-style models fine-tuned from Qwen3.5 and Gemma that return typed decision probabilities in one pass.
ModelbetaPythonApache-2.0 395M ModernBERT System One model serving calibrated Choice, Noul and Score locally as a drop-in /v1/systemone replacement.
ServicebetaPython, TypeScriptGemma 3 4B option scorer that prefills context once and scores all options in one pass, on MLX or PyTorch, with a Doom demo.
ServiceprototypePython- NanoJev★ 2,543
0.6B open Jev replica returning full option distributions with no decoding; one checkpoint plays ViZDoom, Maze and Snake.
ModelresearchPython - jev-visual★ 317
Educational MLX project scoring typed questions about one image with Qwen3.5-0.8B on Apple Silicon, with game demos.
ModelresearchPython