Knowledgator Opens GLiFormer, an Apache 2.0 Schema-Conditioned Extraction Encoder

Knowledgator open-sourced GLiFormer in mid-September 2026 as an Apache 2.0 multi-task information-extraction framework that accepts entity types, class labels, relation types, and nested schemas at inference time without generating free-form tokens. Hugging Face lists knowledgator/gliformer-large-v1 as a 575.6M-parameter DeBERTa-backed checkpoint under license:apache-2.0, and PyPI published gliformer 0.1.2 on September 15, 2026.

Install is pip install gliformer. The Large card documents predict_entities, classify, joint relation inference, structure / Pydantic nested records, multi-task inference_from_schema, and embed_text. A Base v1 checkpoint (264.2M) is also listed. The framework repository is Apache 2.0 and builds on GLiNER-style shared encoders with configurable task heads.

Reported evaluation and scope

The Large model card reports author-run metrics including mean dataset strict NER F1 of 50.91 across 26 datasets, mean classification macro-F1 of 75.03 across 13 datasets, and 91.10 order-free JSON F1 on 500 multilevel structuring examples. Those numbers are project-reported; reproduction scripts live under gliformer_eval in the repo. The card states English text-task coverage and lists limitations around schema wording, span errors, and missing dedicated vision/audio heads on this checkpoint.

Primary sources are the Hugging Face Large and Base model cards, the Apache 2.0 Knowledgator/GLiFormer repository LICENSE, and the PyPI gliformer release history.