Search Papers | Poster Sessions | All Posters
Poster A86 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Roe: Reducing hallucination in large language models via a computational-efficient fine-tuning inspired by human learning and child development
Xueqing Liu1 (), Paul Sajda1; 1Columbia University
Large Language Models (LLMs) have revolutionized Natural Language Processing (NLP) by generating coherent and fluent text for various tasks. However, hallucination remains a significant challenge where LLMs generate entirely fabricated information. This paper introduces a novel approach, "Roe", to counteract hallucination in LLMs. Inspired by psychological studies on promoting honesty in children, we propose a dual-notch objective function guarded by a float loss bar that incentivizes accurate answers and acknowledges uncertainty. This approach improves model honest and enhances accuracy by efficiently fine-tuning the Llama base model, surpassing benchmarks set by models trained on much larger datasets. GPT4All, GPT-3.5, GPT-3, Llama-7B, and Alpaca-7B, exhibit accuracies below 90%, with some even falling below 40%. Our approach achieves comparable accuracy to the state-of-the-art model, GPT4All, while utilizing only a fraction (approximately 1/15) of its training data. When both models are fine-tuned on the same dataset, our method outperforms GPT4All, achieving accuracy rates of nearly or above 95% for all question test sets and 99.32% for the Truthful QA metrics.