Benchmarks & research
Open-Jev provides an open-weight decision model stack using a trained LoRA and scalar head on Qwen to return typed probabilities directly without autoregressive generation.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The 27B v1.1 checkpoint scores 85.28% on the public JevBench subset and needs the pinned upstream Qwen weights plus a custom loader. Prefix caching is experimental and switched off by default. This independent implementation is separate from the other projects named OpenJev.
The released adapter contains changes to a base model, not a complete model on its own. To run Open-Jev locally, pair it with the exact base-model version the maker names and use the custom loader. It then scores the allowed answers without writing a reply.