We’ve just published a new article introducing a Russian-language severity classifier, along with improved English and Chinese models for vulnerability descriptions.
New Russian Severity Classifier and Improved Multilingual Models – Vulnerability-Lookup
These models are trained with VulnTrain and served via ML-Gateway, and are fully integrated into Vulnerability-Lookup.
What’s new
Russian severity classifier based on FSTEC (BDU) data, using a ruRoBERTa-large model
Improved English model trained on diverse sources (CVE, GitHub, PySec, CSAF…)
Improved Chinese model leveraging CNVD data
Better training insights with per-class precision / recall / F1 metrics
Multi-source datasets with traceable origins and dynamic dataset cards
All datasets and models are openly available on Hugging Face.
Under the hood
- VulnTrain 3.1.0 adds FSTEC support, dataset traceability, and improved model selection
- ML-Gateway 0.5.0 now supports multilingual severity classification out of the box
This work is part of the AIPITCH project, supporting practical AI integration in cybersecurity workflows.