Talks - Jodie Burchell: Lies, damned lies and large language models

LLMs are essentially doing “lossy compression” of their training data, leading to information loss and potential hallucinations

Two main types of hallucinations:

• Faithfulness hallucinations: model deviates from given context
• Factuality hallucinations: model generates incorrect facts

Common data quality issues contributing to hallucinations:

• Training data containing misinformation and conspiracy theories
• Low quality sources
• Inadequately filtered web content
• Outdated information

Key methods to reduce hallucinations:

• Retrieval Augmented Generation (RAG)
• Better prompt engineering
• Domain-specific datasets
• Self-refinement and collaborative refinement
• Improved data filtering

RAG implementation considerations:

• Document chunk size
• Choice of embedding model
• Retrieval method
• Vector database selection
• Prompt construction

Model size trends:

• GPT-1: 120 million parameters
• GPT-3: 175 billion parameters
• GPT-4: 1 trillion parameters
• Larger models can encode more information but remain prone to hallucinations

Measuring hallucination rates:

• TruthfulQA dataset for factuality
• HaluEvalQA for faithfulness
• Multiple choice vs. open-ended evaluation
• Need for domain-specific evaluation metrics