Mitigating LLM Hallucination Risk Through Research Backed Metrics

In the context of LLMs, “hallucination” refers to a phenomenon where the model generates incorrect, nonsensical, or unreal text. Identifying and mitigating hallucinations is critical for trustworthy LLM application deployment at scale. In this talk, we will showcase ChainPoll – a unique and powerful methodology to evaluate the quality of LLM outputs, focusing on RAG and fine-tuning use cases. ChainPoll-based metrics have showcased a roughly 85% correlation with human feedback while being low-cost and low-latency to compute. Expected takeaways:

• Deep dive into research-backed metrics to evaluate the quality of the inputs (data quality, RAG context quality, etc.) and outputs (hallucinations) while building LLM-powered applications
• Evaluation and experimentation framework while prompt engineering with RAG and fine-tuning with your own data
• A demo-led practical guide to building guardrails and mitigating hallucinations