> We test our technique in one of the most difficult and nuanced guardrails: the identification of health advice in LLM output, and demonstrate improvement versus other solutions. Our detector is able to outperform GPT-4o by up to 3.73%, despite having 400x less parameters.
4 months agoby mentalgear
27
1