The jev-schema-scorer-deberta-v3-large model has been released for schema classification

Developers introduced jev-schema-scorer-deberta-v3-large — a specialized model based on the DeBERTa-v3-large architecture, designed for data classification with schema consideration. The tool evaluates "state, question, and candidate" pairs to output choice primitives, absence of answer, or evaluation.

The model underwent fine-tuning on a synthetic data array. During testing on version v2, the system demonstrated a choice accuracy of 84.1%. The training and validation process took 1 hour, which allowed achieving a rating of 4 stars according to the project's internal efficiency metric.

The architecture is focused on tasks that require strict binding to data structure, which is critical for automated query processing systems. Using DeBERTa-v3-large as the basis provides high accuracy of context recognition when working with unstructured "question-answer" pairs.

The release is available on the Hugging Face platform and is intended for integration into pipelines that require classification of complex schematic dependencies. The model allows automating candidate filtering in tasks where it is necessary to choose the most suitable option from a proposed data set.