Benchmarking Llama2, Mistral, Gemma and GPT for Factuality, Toxicity, Bias and Propensity for Hallucinations

TL;DR

This paper evaluates the safety, factuality, toxicity, bias, and hallucination propensity of Llama2, Mistral, Gemma, and GPT using fourteen new datasets, highlighting strengths and weaknesses in various safety aspects across models.

Contribution

Introduces fourteen novel safety evaluation datasets and a method for assessing large language models' safety in enterprise tasks, comparing open-source models with GPT.

Findings

GPT outperforms others in safety and factuality

Mistral hallucinates the least but struggles with toxicity

Open-source models' safety degrades in multi-turn conversations

Abstract

This paper introduces fourteen novel datasets for the evaluation of Large Language Models' safety in the context of enterprise tasks. A method was devised to evaluate a model's safety, as determined by its ability to follow instructions and output factual, unbiased, grounded, and appropriate content. In this research, we used OpenAI GPT as point of comparison since it excels at all levels of safety. On the open-source side, for smaller models, Meta Llama2 performs well at factuality and toxicity but has the highest propensity for hallucination. Mistral hallucinates the least but cannot handle toxicity well. It performs well in a dataset mixing several tasks and safety vectors in a narrow vertical domain. Gemma, the newly introduced open-source model based on Google Gemini, is generally balanced but trailing behind. When engaging in back-and-forth conversation (multi-turn prompts), we find that the safety of open-source models degrades significantly. Aside from OpenAI's GPT, Mistral is the only model that still performed well in multi-turn tests.