metask-jev

A research project (documented in Chinese) exploring open-weight typed-decision models in the Jev class, taking a state plus a bounded rubric and returning calibrated probabilities per option in a single forward pass with no generation or parsing; its metask-jev-4b model is reported to outperform both Bespoke's Nimble-9B and Jev itself on the JevBench benchmark.

unverified — sourced from awesomejev.com; description independently written by AionEdge from the repo’s own GitHub metadata and README, not GitHub-API relevance-checked