Reducing Hallucinations and Evaluating LLMs for Production

Explore the challenges of evaluating Large Language Models (LLMs) and reducing hallucinations in their outputs in this informative conference talk. Gain insights into traditional evaluation methods like BLEU and F1 scores, as well as modern approaches such as Eleuther's AI evaluation framework. Examine the underlying causes of LLM hallucinations, including biases in training data and overfitting. Learn about open-source LLM validation modules and best practices for minimizing hallucinations to prepare LLMs for production use. Designed for practitioners, researchers, and enthusiasts with a basic understanding of language models, this talk provides valuable knowledge on improving LLM performance and reliability.

Reducing Hallucinations and Evaluating LLMs for Production - Divyansh Chaurasia, Deepchecks