Calibrated Typed Decisions for Entity, Relation, and Event Extraction
Calibrated Typed Decisions for Entity, Relation, and Event Extraction
CalOn 是一款 0.6B 模型,面向資訊抽取中的快速決策,可用指定類型處理實體、關係與事件。在 NER 基準測試中,CalOn 每秒處理 120 句,吞吐量是 GLiNER2 的 5.5×、GLiFormer 的 6.7×;評估涵蓋 40 個基準。模型權重、推理程式碼及訓練程式碼以 Apache 2.0 授權開源。
Today we’re releasing CalOn, our 0.6B model for fast decisions in information extraction. ⚡
In our NER benchmark, CalOn processes 120 sentences per second. That’s 5.5× the throughput of GLiNER2 and 6.7× that of GLiFormer.
We built on TypeSafe’s RLCD and adapted the training approach to spans and span pairs. The same model handles entities, relations, and events using the types you specify.
Our evaluation spans 40 benchmarks. Beyond standard NER, RE and EE, we also evaluate aspect sentiment triplet extraction, pronoun resolution, entity and event coreference, and temporal, causal and subevent relations between events.
We express these tasks as schemas over spans and span pairs and run inference with the same checkpoint. CalOn leads the evaluated baselines on these converted benchmarks.
The weights, inference code, and training code are open under Apache 2.0.
🤗 https://huggingface.co/calon-ai/calon-0.6b
💻 https://github.com/extraordinarylab/calon
📄 https://arxiv.org/abs/2610.17113
來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co