FLAME: A Lightweight Language Model for Spreadsheet Formulas

Large language models are expensive to train and deploy for assisting with Excel formula authoring. This paper introduces FLAME, a transformer-based model trained exclusively on Excel formulas. With only 60 million parameters and a fraction of the training data used by larger models, FLAME achieves competitive or even superior performance on formula repair, completion, and retrieval tasks compared to models like Codex and CodeT5. This is attributed to its novel pre-training objectives and Excel-specific tokenizer.