Towards Efficient Data Wrangling with LLMs using Code Generation
Xue Li (UvA/MotherDuck), Till Döhmen (MotherDuck)
Abstract
While LLM-based data wrangling approaches that process each row of data have shown promising benchmark results, computational costs still limit their suitability for real-world use cases on large datasets. We revisit code generation using LLMs for various data wrangling tasks, which show promising results particularly for data transformation tasks (up to 37.2 points improvement on F1 score) at much lower computational costs. We furthermore identify shortcomings of code generation methods especially for semantically challenging tasks, and consequently propose an approach that combines program generation with a routing mechanism using LLMs.