Lies, Damned Lies, and Large Language Models - Measuring and Reducing Hallucinations

Explore the challenges and solutions surrounding large language models' (LLMs) tendency to produce incorrect information or "hallucinate" in this 29-minute conference talk from EuroPython 2024. Delve into the main causes of hallucinations in LLMs and learn how to measure specific types of misinformation using the TruthfulQA dataset. Discover practical techniques for assessing hallucination rates and comparing different models using Python tools like Hugging Face's `datasets` and `transformers` packages, as well as the `langchain` package. Gain insights into recent initiatives aimed at reducing hallucinations, with a focus on retrieval augmented generation (RAG) and its potential to enhance the reliability and usability of LLMs across various contexts.