Oryn, an ultra fast compact decision model (huggingface.co)

Oryn is an innovative compact decision model designed for low-latency inference, built on the boltuix/bert-mini architecture. Unlike traditional classifiers that rely on a fixed output layer, Oryn dynamically receives classification questions and candidate options at inference time, enabling it to tackle a variety of decision-making tasks such as binary classification, multiclass classification, and ordinal scoring. Weighing in at approximately 21 MB, it boasts a warm-up inference time of about 3 ms for single requests in local testing, marking it as suitable for applications requiring rapid responses, such as phishing detection and message triage. The significance of Oryn lies in its versatile architecture, which allows for the integration of different question types and scoring options. The model's performance metrics reveal an overall accuracy of 71.7%, excelling in binary and categorical tasks, although it demonstrates lower reliability in ordinal sentiment classification, where it achieves an accuracy of only 25.82%. This makes Oryn a promising tool for real-time applications and dynamic classification research, particularly in environments where low-latency and local inference are critical. Its adaptable nature positions it as a valuable asset for developers experimenting with a range of classification scenarios.