Uncertainty quantification in fine-tuned LLMs using LoRA ensembles

TL;DR

This paper introduces a method for quantifying uncertainty in fine-tuned large language models using low-rank adaptation ensembles, providing insights into model knowledge retention and trustworthiness.

Contribution

It presents a novel approach combining posterior approximations with low-rank adaptation ensembles for uncertainty quantification in fine-tuned LLMs.

Findings

Enables principled uncertainty estimation in fine-tuned LLMs.

Reveals retention of prior knowledge during overfitting.

Analyzes model behavior across multiple-choice datasets.

Abstract

Fine-tuning large language models can improve task specific performance, although a general understanding of what the fine-tuned model has learned, forgotten and how to trust its predictions is still missing. We derive principled uncertainty quantification for fine-tuned LLMs with posterior approximations using computationally efficient low-rank adaptation ensembles. We analyze three common multiple-choice datasets using low-rank adaptation ensembles based on Mistral-7b, and draw quantitative and qualitative conclusions on their perceived complexity and balance between retained prior knowledge and domain specific adaptation during and after fine-tuning. We identify unexpected retention of acquired knowledge during fine-tuning in the overfitting regime.