Noise-Aware Fine-Tuning of LLMs for Robust Text Classification

Main Article Content

Ramakrishnan Sathyavageeswaran

Abstract

Large Language Models (LLMs) have shown exceptional ability and success in a broad set of natural language processing (NLP) tasks, especially in text classification. Their ability, however, drops considerably in the face of noisy inputs which are typical of realworld scenarios, e.g. social media content, OCR-scanned documents, or speech-to-text output. The aim of the research paper is to increase the robustness of LLMs to different forms of noises, such as lexical, syntactic, semantic, and label noise, by using the noiseaware fine-tuning methods. We also compare several strategies that are robustnessfocused, such as noise-augmented fine-tuning, adversarial training and contrastive learning, in terms of their inference performance on a variety of architectures and data sets. The fact that QLoRA can be used with 13B+ parameter models enables us to achieve efficiency via parameter-efficient fine-tuning (PEFT). As we show, there are substantial performance gains to be made by training noise sensitivity, especially in noisy scenarios, and only by moderately reduced trade-offs in clean-data accuracy (e.g., + 6.1 percentage points gain at 30 percent lexical noise, - 0.3 percentage points on clean data). We also suggest a feasible noise typology, describe an effective training model and provide deployment suggestions. Such contributions reduce the difference between benchmark performance and practical reliability, such that LLMs may be able to perform better when exposed to unstructured and imperfect data.

Article Details

Section
Articles