Exploration of LLM Lossless Compression on Scientific Data

Recently, large language models (LLMs) have demonstrated significant potential across various scientific fields, such as microscopy, medical imaging, and protein structure prediction. LLM-powered compression techniques, particularly those designed for lossless compression, have shown promise by delivering superior quality and higher compression ratios compared to traditional methods. Despite these advancements, LLMs often face challenges in effectively managing and processing scientific data due to the complexity of the compression context. Compression context is loss in dimensionality information that occurs when higher dimensional data is transformed down to fit into compressors. Consequently, LLM-based compression tends to underperforms on scientific datasets. To address this issue, this work conducts a comparative analysis of LLM-based compression and general compression methods to gain a deeper understanding of their effectiveness on scientific data. We propose a fine-tuning optimization approach tailored for an LLM compression model, aiming to enhance its compression percentage. The evaluation results reveal that LLM-based compression outperforms other compression methods on compression percentage by up to 36.1%, including eight state-of-the-art compressors, on STEM, CLIC, RTM, KODAK, and COVID datasets. This demonstrates the potential of specialized LLMs for effective compression of scientific data.