Towards Better Serialization of Tabular Data for Few-shot Classification with Large Language Models
This work addresses the challenge of efficiently processing tabular data for few-shot classification tasks using LLMs, representing an incremental improvement over existing methods like TabLLM.
The paper tackles the problem of tabular data classification with large language models by introducing three novel serialization techniques, with the LaTeX method boosting performance on domain-specific datasets while maintaining memory efficiency.
We present a study on the integration of Large Language Models (LLMs) in tabular data classification, emphasizing an efficient framework. Building upon existing work done in TabLLM (arXiv:2210.10723), we introduce three novel serialization techniques, including the standout LaTeX serialization method. This method significantly boosts the performance of LLMs in processing domain-specific datasets, Our method stands out for its memory efficiency and ability to fully utilize complex data structures. Through extensive experimentation, including various serialization approaches like feature combination and importance, we demonstrate our work's superiority in accuracy and efficiency over traditional models.