AnyJev gives an open model the same typed-decision output Jev gives natively: a choice, a yes/no, a score, or a probability, without any fine-tuning.
It works at three levels depending on how much labelled data you have. L0 needs none and just reduces position bias in the raw readout. L1 uses 100 to 500 labelled examples to calibrate the probabilities. L2 uses 100 to 300 labels to fit a small closed-form head on a hidden state, which can even stop the forward pass early once the decision is in.
Because L2 only needs a hidden state, the authors show it can run behind an ordinary vLLM embedding server: three commands turn a model on the Hub into a calibrated decision endpoint. Built by researchers at Nokia and Tencent Hunyuan.
What you get
- Full Python source code under Apache 2.0
- The L0/L1/L2 deciders and the vLLM serving pipeline
Jev
Python
LLM
Codex
Claude Code
