yinsongxu/llm2jev is a Python project that turns an existing language model into a structured judgment model in the Jev style. According to the repository description, it can be used without training or fine-tuning. The latest release v0.2.0 adds Box judgment for image bounding boxes. Four query types, including the existing Choice, Score and Noul, can be mixed in a single request.
Instead of generating answer tokens one by one, it reads only the yes/no logits during the prefill stage that processes the input, obtaining the judgment probabilities. The point is therefore to apply it to classification, evaluation and action selection without token-level decoding. It handles text and image inputs, and offers SGLang, Transformers and MLX backends. Its client inherits the OpenAI Python SDK, so chat and judgment can be handled on the same service and model. The license is Apache-2.0.
The README evaluates accuracy and latency with the public JevBench 231 questions and claims that P50 latency is less than one tenth of Jev's official results. However, this comparison is the repository's own evaluation, and no external reproduction is given in FACTS. So rather than taking its performance advantage as independently confirmed, it is more appropriate to see this as a project that demonstrates a prefill-based judgment approach and the Box addition.