Lecture Notes on Detecting Hallucinations in LLM Applications

Introduction to the Speaker and Topic

• Speaker: Alex Thomas, Principal Data Scientist at y Cube.
• Topic: Detecting hallucinations in AI and measurement systems for detection.

Background on LLMs and Hallucination

• LLMs: Large Language Models used for processing natural language with significant advancements.
• Problem: LLMs can hallucinate, i.e., generate false information which can:
  • Spread misinformation.
  • Erode trust and brand reputation.
  • Lead to legal and customer trust issues.
• Causes: Bias inherent in LLMs, user-framing issues.

Examples of Hallucinations Impact

• Example: Air Canada's chatbot invented a refund policy.
• Impact: Legal and brand reputation damage.

Measuring Hallucinations

• Terminology:
  • Application: LLM-based application (e.g., chatbot, summarization system).
  • Measurement System: System for measuring the LLM application.

Types of LLM Tasks

1. Zero Context

   • Basic chatbot experience with small inputs/outputs.
   • Challenges: External data source needed for factual accuracy.
   • Susceptible to user framing.

2. RAG Q&A (Retrieval Augmented Generation)

   • Combines user input with retrieved text for grounded answers.
   • Challenges: Multiple sources of error (LLM hallucination, incorrect retrieved data).

3. Summarization

   • Large input, smaller output.
   • Complex accuracy concept.
   • Challenges: Maintaining factual integrity when summarizing potentially false articles.

Establishing Benchmarks and Metrics

• Challenges:
  • Lack of established metrics or conventions.
  • Focusing on factual accuracy over text fluency.
• Measurement Strategies:
  • Pythia Strategy: Extracts claims as triples for evaluation.
  • Grading Strategy: Simple letter grade evaluation.
  • Lynx Strategy: Pass/fail evaluation focused on faithfulness.

Models Tested

• Models: GPT-4.0 and LLaMA 3.1 (70 billion parameters).

Measuring Measurement Systems

• Comparison with NLP Benchmarking:
  • Old benchmarks specialized by tasks/domains.
  • New LLM benchmarks murkier, less established.

Data Sets for Measuring

• Summarization Data Sets:
  • CNN/Daily Mail, BBC articles.
  • Expert vs. Mechanical Turker labels.
• RAG Q&A Data Set:
  • Based on binary pass/fail labels.

Metrics for Evaluation

• Spearman Correlation: Issues with low granularity labels.
• Mean Absolute Error: Better for summarization evaluation.
• Accuracy: Used for binary classification in RAG Q&A.

Conclusion

• Complexity: Measuring LLM applications requires sophisticated systems.
• Understanding: Need to understand data, metrics, and system tuning.
• Future Work: Testing more strategies, models, and exploring calibration.

Questions and Recommendations

• Data Comparison: Compare distribution of open data sets with your own.
• Industry-Specific Use: Pythia can be adapted for various industries, though specialization is beneficial.

Final Notes

• Resources: Links available for trying out systems.
• Webinar Continuation: Stay tuned for future developments.