Dr. LLaMA: Improving Small Language Models in Domain-Specific QAvia Generative Data Augmentation

https://arxiv.org/pdf/2305.07804.pdfhttps://arxiv.org/pdf/2305.07804.pdf

Our findings indicate that LLMs effectively refine and diversify existing question-answer pairs, resulting in improved performance of a much smaller model on domain-specific QA datasets after fine-tuning.This study highlights the challenges of using LLMs for domain-specific question answering and suggests potential research directions to address these limitations, ultimately aiming to create more efficient and capable models for specialized applications.

Fine-tuning Large Language Models (LLMs) for specific tasks poses computational and time-related challenges (Liu et al., 2022; Vos et al., 2022). To address these issues, researchers have developed efficient fine-tuning techniques, such as Prefix Tuning and Low-rank Adaptation, as alternatives to traditional fine-tuning methods.

 

Generative data augmentation is a vital technique in machine learning for expanding and diversifying training data, ultimately enhancing model generalization (Calimeri et al., 2017; Shao et al., 2019; Sandfort et al., 2019; Shin et al., 2018; Yang et al., 2020; Carlini et al., 2021).

For NLP task, generative data augmentation with LLMs can involve paraphrasing text, creating alternative question-answer pairs, or generating new sentences or paragraphs. Producing diverse representations of input data enables models to learn various ways to express the same underlying concepts, increasing their adaptability to real-world data variations.

However, ensuring the quality and relevance of generated samples is crucial, as low-quality or irrelevant data can negatively impact performance. Additionally, controlling the diversity of generated samples is essential to prevent redundancy or overly similar data points. Thus, generative data augmentation using LLMs in NLP holds promise for improving model generalization and performance while addressing data quality, relevance, and diversity challenges.

Instruction-tuning constrains domain adaptability of language models