This repo is an open reproduction of TypeSafe AI's Jev model, providing a 150M-parameter non-autoregressive typed decision engine. It scores 0.697 accuracy on the typed-decisions benchmark (Jev: 0.727), with better calibration, and can be trained for free on Colab.
该仓库明确复现 TypeSafe AI 的 Jev 模型，包含完整代码、训练与评估流程，并在 typed-decisions 基准上与 Jev 对比，属于深入研究/复现项目。
JevalsJevals/jevals-dataIndependent benchmark data comparing TypeSafe AI's Jev (System One model) against LLMs on decision score, accuracy, calibration, cost, and latency for typed tasks.
NearCaiNearCai/JevLightA CityFlow-based traffic signal control framework that evaluates TypeSafe Jev and OpenAI-compatible LLM controllers alongside rule-based and reinforcement-learning baselines through the same simulation pipeline.
choxoschoxos/jev-reviewerUses TypeSafe AI's Jev/System One model (jev-1.13.0) to locate verbatim evidence in trial reports and supplements, supporting RoB 2, ROBINS-I, QUADAS-2 templates; files are processed locally in the browser.
