Kotoba Labs released the open-jev-deberta-v3-large model

Team Kotoba Labs introduced the open language model open-jev-deberta-v3-large, created for structured decision-making tasks. The novelty provides calibrated probability distributions for various types of questions in one pass.

The architecture is built on the basis of the DeBERTa-v3-large base and optimized for processing options, ratings, and uncertain questions within a single forward computation. This approach reduces computational costs when inferring complex analytical pipelines.

In specialized tests, the model demonstrated an accuracy of 85.4% on in-domain data. In out-of-domain testing (OOD) conditions, this figure is 69%, which confirms the algorithm's robustness to new types of input data.

The weights and documentation are published on the Hugging Face platform under a free license. The release is aimed at developers of automated analytics systems that require high accuracy of classification and risk assessment.