Diego Carpintero

Well done! Just sharing PangolinGuard, a fine-tuned model from ModernBERT-large aimed at implementing AI guardrails. Despite its small size, the model closely approximates the performance of Claude 3.7 and Gemini Flash 2.0 on a mixed benchmark (based on BIPIA, NotInject, Wildguard-Benign and PINT). I believe this could provide a lightweight, inexpensive approach for (i) adding custom, self-hosted safety checks, (ii) steering conversations to compliant topics, and (iii) mitigating risks when connecting AI pipelines to external services.

If someone wants to explore the use of ModernBERT for AI safety, feel free to get in touch:
📝 article | 🤗 hf-space | repo