Developers released the GLiNER2.5 Base model for information extraction

The development team introduced GLiNER2.5 Base — a new version of the model designed for unified extraction of structured data from unstructured text. The release expands the line of natural language processing tools, offering a balance between performance and resource intensity.

The model is built on the DeBERTa-v3-base architecture and contains 194 million parameters. The system's main specialization is entity extraction, object classification, and identification of relationships between them. To work with data, the model implements support for the AutoExtractor functionality, which allows automating the process of forming extraction schemes.

The technical implementation of the model is focused on tasks that require high accuracy when working with arbitrary types of entities that were not rigidly defined during training. The use of DeBERTa-v3 provides improved context understanding compared to previous iterations of the GLiNER family, which is critical for relationship extraction tasks in complex sentences.

The update complements the ecosystem of tools for automating text array analysis. The developers note that this version is optimized for integration into data processing pipelines, where flexible adjustment of extraction schemes is required without the need to retrain the model for each specific domain or entity type.