User repository on GitHub
- Forks
- 121
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Repos & SDKs
Jevlike
Train 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.
- Stars
- 1.4k
- Language
- Python
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Last verified
- Sep 2026
About this project
Train a small model that chooses among a changing list of text options. A Jev-like model takes a piece of text and a list of N text options. It returns one probability for each option. It does this in one pass instead of writing an answer word by word. Jev is TypeSafe's commercial model for this kind of task. TypeSafe has not published its design. This repository is an independent starter model with the same input and output shape.
Quoted from the project's README. Read the full README.
Works forJev for Realtime control
Used inJev for Gaming
Highlights
- Training rows are JSONL with context, options, and a zero-based label; each row may have a different number of options, minimum two.
- The default byte encoder truncates context to 192 bytes and each option to 32 bytes; --context-tokens and --option-tokens raise that.
- On target-disjoint Wikispeedia, a frozen Qwen2.5-0.5B encoder plus head reached 26% next-click vs about 8% for shuffled controls.
- The Doom joint checkpoint averaged 0.60 kills and -97.50 reward over ten episodes; the chess head went 0W/2D/48L vs Stockfish level 0.
- At eight options, one pass was about 100 times faster than a small decoder forced to write 400 tokens.
Quickstart
bash
uv venv
source .venv/bin/activate
uv pip install -e '.[dev]'
jevlike-data synthetic --output data/synthetic
jevlike-train data/synthetic/train.jsonl --validation data/synthetic/validation.jsonl --output runs/synthetic.pt
jevlike-eval runs/synthetic.pt data/synthetic/test.jsonlWatch out
It is an independent research starter with the same input and output shape as Jev, not a copy, and the repo says it did not reproduce TypeSafe's private training method or equal quality.
Questions about this entry
Short answers drawn from the linked project or page.
Just the shape. TypeSafe has not published Jev's design, and the repo states it did not show equal quality with Jev or reproduce TypeSafe's private training method. It is a research starter: context plus a list of options in, one probability per option out.
You need the complete option list before you can predict, so it does not help when candidates are generated one at a time. The default byte encoder also truncates context to 192 bytes and each option to 32, and is described as cheap but weak on language meaning.
Mixed. About 98% comes from synthetic menus the project generates. On target-disjoint Wikispeedia next-click data, a frozen Qwen2.5-0.5B encoder plus the head reached 26% against roughly 8% for shuffled and random-encoder controls; a small model trained from scratch on 40,000 clicks got 29%.
It is against a small local decoder forced to write 400 tokens at eight options, and the README says so rather than claiming parity with a large commercial model. The same honesty shows in the games: the chess head went 0 wins, 2 draws, 48 losses to Stockfish level 0.
Install with uv, then jevlike-data makes a synthetic set, jevlike-train fits the scorer, jevlike-eval scores a held-out file, and jevlike-predict scores one new menu. --device selects CPU, Apple MPS, or CUDA. The default encoder learns byte embeddings from scratch; --encoder hf freezes a Hugging Face encoder instead, and the checkpoint stores the trained head and the encoder name rather than the encoder weights.
Keep related records together, for example every record for one customer or one target page in a single split, so near-duplicates cannot leak into the test set, and run eval once on a held-out file never used for training or model selection. The report prints top-1 accuracy, top-3 accuracy, expected calibration error, and a shuffled-context control that pairs each menu with the wrong context.
The code is MIT. Downloaded datasets and pretrained models keep their own terms: the Wikispeedia build script pulls the public SNAP archives and keeps the data outside the repository, and the source data are Robert West and Jure Leskovec's Human Wayfinding in Information Networks, WWW 2012, with terms on the SNAP dataset page.
More Jev resources like this
A small, mechanics-free System One decision core for game NPCs: the adapter supplies an observation and opaque legal actions, and a scripted or Jev provider picks one, with a deterministic offline provider for tests.
Jev Repos & SDKs#community#github#python
Community
safiullahrattar/sys1npcPython
safiullahrattar/sys1npcPythonAn open 0.6B replica of Jev that turns states and questions into full probability distributions without decoding answer tokens, trained and evaluated on Maze, Snake, and ViZDoom.
Jev Repos & SDKs#community#python#open-models
Community
TianyuCodings/NanoJevPython
TianyuCodings/NanoJevPythonNokia Applied Research's package that turns any open LLM into a Jev-style decision model: typed questions are read from a single prefill, and the calibration fixes need no fine-tuning.
Jev Repos & SDKs#community#python#open-models
Community
nokia-applied-research/AnyJevPython
nokia-applied-research/AnyJevPythonAn open, non-autoregressive System One model with published weights and a TypeSafe-compatible /v1/systemone API; the 395M-parameter OptionMarker checkpoint scores discrete and ordinal questions locally.
Jev Repos & SDKs#community#python#open-models
Community
wfzyx/vonPython
wfzyx/vonPythonTogether AI's open recipe and weights for a Jev-inspired decision model fine-tuned on Qwen3.5-4B, with the full data pipeline, training config, and saved benchmark results.
Jev Repos & SDKs#community#python#open-models
Community
togethercomputer/tev1Python
togethercomputer/tev1PythonOpen infrastructure from the SimpleJev project for training, evaluating and deploying System 1 decision models across language and multimodal backbones: one API takes a state, question and candidate options and returns a choice with probabilities.
Jev Repos & SDKs#community#github#python
Community
SimpleJev/JevAnyPython
SimpleJev/JevAnyPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.