Fastino Releases GLiNER2.5 Boundary Architecture
Fastino releases three GLiNER2.5 checkpoints on Hugging Face under Apache 2.0, removing span enumeration to improve information extraction.
Stock photo for illustration only, not from the actual event
- Fastino released three GLiNER2.5 checkpoints on Hugging Face under the Apache 2.0 license.
- The new architecture removes span enumeration to maintain linear computation relative to sequence length.
- GLiNER2.5 Base achieves an average score of 54.87, improving from the previous 53.34.
- The XNLI benchmark for the Multi model surges to 62.30, marking a 24.75-point increase.
Fastino has officially released three GLiNER2.5 checkpoints on Hugging Face under the Apache 2.0 license, enabling local inference on CPU, CUDA, or MPS via pip installation for Python 3.10 and above. Since no inference providers currently host these model checkpoints, self-hosting remains the primary deployment pathway for developers looking to integrate the system.
Previous iterations of GLiNER located entities by exhaustively enumerating candidate spans, pairing every start position with every permitted width, and scoring each against the target schema. That architectural design tightly coupled computation to a width axis, imposing a hard ceiling on the maximum entity length the system could process.
Stock photo for illustration only, not from the actual event
In GLiNER2.5, Fastino has completely removed span enumeration. The shared encoder processes text and schema queries in a single pass. Instead of scoring individual spans, the model predicts start and end scores across token boundaries alongside inside token scores. A sparse proposal stage subsequently selects the most promising starts and ends per query, pairing them without imposing distance restrictions.
A newly introduced reranking head evaluates each candidate leveraging boundary evidence and span content, while relation candidates are drawn from the same pool rather than requiring a separate processing path. The Fastino team reports that computational complexity remains strictly linear with respect to sequence length under a fixed schema and candidate budget.
Shifting away from traditional span enumeration to a boundary prediction paradigm addresses a major computational bottleneck in information extraction pipelines. By eliminating exhaustive combinatorial checks, the model handles varying text lengths much more efficiently, paving the way for scalable and rapid unstructured data parsing in real-world applications.
All three new models share a unified public API and must be loaded using AutoExtractor rather than legacy span loaders. In zero-shot evaluations across 16 public datasets comparing Macro F1 against GLiNER2 at matched sizes, GLiNER2.5 Multi reached an overall average of 56.17 compared to 56.09, while the Base model improved to 54.87 from 53.34, alongside notable gains in untrained languages like Romanian RONEC.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment