- Published on
Knowledgator Releases GLiFormer: A 575M-Parameter Encoder That Hits 91.10 F1 on Nested JSON Extraction Without Genera...
- Authors
- Name
- DeepHorus
Knowledgator Engineering has unveiled GLiFormer, a schema-conditioned encoder framework that achieves a state-of-the-art 91.10 F1 score on nested JSON extraction without generating any tokens. The release includes two checkpoints on Hugging Face: GLiFormer Base v1 with 264.2 million parameters and GLiFormer Large v1 with 575.6 million parameters. Both are available under the Apache 2.0 license, can be installed via pip install gliformer, and run on either CPU or GPU. This marks a significant shift from traditional generative approaches to information extraction.
The problem GLiFormer targets is the fragmented nature of extraction pipelines. Typically, developers chain separate models: one for named-entity recognition (NER), another for text classification, a third for relation extraction, and yet another to reconstruct records into nested JSON. The research team argues these tasks share a core operation: encode the source text, represent the requested concepts, and score their compatibility. While large language models can emit nested JSON, they do so token by token, generating field names, punctuation, and values. GLiFormer eliminates output generation from that path entirely, offering a more deterministic and efficient alternative.
At its core, GLiFormer builds on GLiNER and generalizes label matching through an "anchor." An anchor is the object each runtime label is scored against—it can be a group vector for classification, an entity pair for relations, or a record slot. The source text is encoded once, and multiple schemas for the same document then run as task-local groups over that shared encoding. For NER, the head scores start, end, and inside evidence for every token and label pair, with independent sigmoid outputs allowing nested mentions and shared boundaries to coexist. Structuring runs in four stages: grounding field values as spans taken directly from the source text, assigning spans to unordered record slots via Hungarian matching, predicting directed parent-child links restricted to schema-allowed paths, and assembling nested JSON with a deterministic decoder. Because values are source spans, the model cannot invent value text missing from the input, though span selection, record assignment, and hierarchy can still be wrong.
Both v1 checkpoints use the gliformer-layout model type with five heads: NER, classification, joint relations, multilevel structuring, and embeddings. Each configures a 12-word maximum span width and 100 record anchors. GLiFormer Base has 12 encoder layers, an embedding dimension of 768, and a configured max length of 16,384 tokens. GLiFormer Large has 24 layers, 1024 dimensions, and a max length of 8,192. Base starts from a DeBERTa backbone further pretrained on 100 billion tokens. The paper documents 1,357,671 examples for broad multitask training and 372,090 for task-focused post-training.
Benchmarks reported by Knowledgator show GLiFormer Large scoring 91.10 F1 on nested JSON extraction across 500 examples, while Base scores 87.20. For comparison, GPT-5.6-luna scores 91.96 and GPT-5-mini 82.56. Notably, the metric is order-free and boundary-tolerant, not exact JSON match, which is a more realistic evaluation for structured extraction. On classification across 13 datasets, Large reaches 75.03 mean macro-F1 and Base 72.36. These results demonstrate that an encoder-only architecture can rival much larger generative models on specific extraction tasks.
The impact of GLiFormer is twofold. First, it simplifies deployment: a single model handles NER, text classification, relation extraction, nested JSON structuring, and text embeddings, reducing the need for multiple specialized models. Second, by eliminating token generation, it avoids common LLM pitfalls such as hallucinated values, punctuation errors, and high inference costs. The open-source release under Apache 2.0 and straightforward pip installation make it immediately accessible to developers. Running on CPU or GPU further broadens its applicability, from edge devices to cloud servers.
Reactions from the AI community are likely to be positive, given the growing demand for efficient, controllable information extraction. The Hugging Face checkpoints provide a low barrier to entry, and the detailed benchmarks will appeal to practitioners comparing against LLM-based solutions. Looking ahead, Knowledgator may expand the framework to support more schema types, improve hierarchical accuracy, and integrate with existing data pipelines. As retrieval-augmented generation and structured data extraction become increasingly critical, GLiFormer's token-free, schema-conditioned approach could set a new standard for production-grade information extraction.
