{"id":118314,"date":"2026-09-17T01:49:45","date_gmt":"2026-09-17T01:49:45","guid":{"rendered":"https:\/\/youzum.net\/knowledgator-releases-gliformer-a-575m-parameter-encoder-that-hits-91-10-f1-on-nested-json-extraction-without-generating-tokens\/"},"modified":"2026-09-17T01:49:45","modified_gmt":"2026-09-17T01:49:45","slug":"knowledgator-releases-gliformer-a-575m-parameter-encoder-that-hits-91-10-f1-on-nested-json-extraction-without-generating-tokens","status":"publish","type":"post","link":"https:\/\/youzum.net\/zh\/knowledgator-releases-gliformer-a-575m-parameter-encoder-that-hits-91-10-f1-on-nested-json-extraction-without-generating-tokens\/","title":{"rendered":"Knowledgator Releases GLiFormer: A 575M-Parameter Encoder That Hits 91.10 F1 on Nested JSON Extraction Without Generating Tokens"},"content":{"rendered":"

Knowledgator Engineering has released GLiFormer<\/a>, a schema-conditioned encoder framework for information extraction. One model handles named-entity recognition (NER), text classification, relation extraction, nested JSON structuring, and text embeddings. You pass labels and extraction schemas at inference time. Two checkpoints are on Hugging Face. GLiFormer Base v1<\/a> has 264.2M parameters, and GLiFormer Large v1<\/a> has 575.6M.<\/p>\n

Deployable today? Yes.<\/strong> Both checkpoints are Apache 2.0, install with pip install gliformer<\/code>, and run on CPU or GPU. <\/p>\n

The Problem It Targets<\/strong><\/h2>\n

Extraction stacks often chain separate models. One tags entities, another classifies documents, and a third rebuilds records. The research team<\/a> argues these tasks share one core operation. Encode the source, represent the requested concepts, then score their compatibility.<\/p>\n

LLMs can emit nested JSON, but they generate field names, punctuation, and values token by token. GLiFormer removes output generation from that path.<\/p>\n

How GLiFormer Works<\/strong><\/h2>\n

GLiFormer builds on GLiNER<\/a> and generalizes its label matching through an \u2018anchor.\u2019 An anchor is the object each runtime label gets scored against. It can be a group vector for classification, an entity pair for relations, or a record slot.<\/p>\n

The source is encoded once. Multiple schemas for the same document then run as task-local groups over that shared encoding. Head compute still grows with the number of groups, labels, and anchors.<\/p>\n

For NER, the head scores start, end, and inside evidence for every token and label pair. Independent sigmoid outputs let nested mentions and shared boundaries coexist.<\/p>\n

Structuring runs in 4 stages:<\/strong><\/p>\n

    \n
  1. Ground field values as spans taken directly from the source text.<\/li>\n
  2. Assign spans to unordered record slots, trained with Hungarian matching.<\/li>\n
  3. Predict directed parent-child links, restricted to paths the schema allows.<\/li>\n
  4. Assemble nested JSON with a deterministic decoder.<\/li>\n<\/ol>\n

    Values are source spans, so the model cannot invent value text missing from the input. Span selection, record assignment, and hierarchy can still be wrong.<\/p>\n

    <\/div>\n

    \n

    Checkpoints and Training<\/strong><\/h2>\n<\/p>

    Both v1 checkpoints use the gliformer-layout<\/code> model type with 5 heads: NER, classification, joint relations, multilevel structuring, and embeddings. Each configures a 12-word maximum span width and 100 record anchors. Full specs sit in the pretrained models docs<\/a>.<\/p>\n

    \n\n\n\n\n\n\n\n\n
    Spec<\/th>\nBase v1<\/th>\nLarge v1<\/th>\n<\/tr>\n<\/thead>\n
    Parameters<\/td>\n264.2M<\/td>\n575.6M<\/td>\n<\/tr>\n
    Encoder layers<\/td>\n12<\/td>\n24<\/td>\n<\/tr>\n
    Embedding dimension<\/td>\n768<\/td>\n1024<\/td>\n<\/tr>\n
    Configured max_len<\/code><\/td>\n16,384<\/td>\n8,192<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

    GLiFormer-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.<\/p>\n

    Benchmarks<\/strong><\/h2>\n

    All scores below are reported by Knowledgator.<\/strong><\/p>\n