README.md

metadata

title: README
emoji: 💻
colorFrom: blue
colorTo: red
sdk: static
pinned: false

Enguard AI

One guardrail for all, all guardrails for one!

Why should you use these models?

• Optimised for precision to reduce false positives.
• Extremely fast inference using static embeddings powered by Model2Vec.

Which guards are available?

Below is an overview of all guardrails, showing the best results for the smallest (-2m), best-performing, and multi-lingual models across all configurations.

Dataset	Classifies	Collection	Smallest (2m)	Best Performing	Multi-lingual (128m)
harmfulness-mix	prompt-harmfulness	Collection	0.9192	0.9350	-
intel	response-politeness	Collection	0.8795	0.8908	0.8908
jailbreak-in-the-wild	prompt-jailbreak	Collection	0.8515	0.8905	0.8905
jailbreak-sok	prompt-jailbreak	Collection	0.9762	0.9810	0.9810
jigsaw	prompt-toxicity	Collection	0.8967	0.9067	0.8986
nvidia-aegis	response-safety	Collection	0.7612	0.7760	0.7530
nvidia-aegis	prompt-safety	Collection	0.7957	0.8131	0.7929
polyguard	response-safety	Collection	0.8635	0.8808	0.8753
polyguard	response-refusal	Collection	0.8952	0.9039	0.9015
polyguard	prompt-safety	Collection	-	0.9331	0.9255
toxic-chat	response-jailbreak	Collection	0.9872	0.9914	-
toxic-chat	prompt-toxicity	Collection	0.9515	0.9555	-