Research available models and libraries for LLM / chatbot guardrailing

[jacopo-chevallard]

The goal is to do a comprehensive, but quick, research of the available approaches (models, libraries, …) that can be used as guardrails for our chatbot, to filter malicious/harmful inputs and outputs.

Useful models (they are all multilingual):

• Meta prompt-guard: https://huggingface.co/meta-llama/Prompt-Guard-86M (prompt injection / jailbreak)
• Toxicity classifier: https://huggingface.co/textdetox/xlmr-large-toxicity-classifier
• Personal information detection: https://huggingface.co/iiiorg/piiranha-v1-detect-personal-information

[linear]

CORE-311 Research available models and libraries for LLM / chatbot guardrailing