LLMs for Understanding and Preserving Historical Musical Codes : Evaluation and Benchmarking

This thesis focuses on the collection, preparation, and structuring of a curated code dataset centered around Csound projects, aimed at enabling future tasks such as code translation and analysis. We detail the process of selecting clean, well-documented repositories, preprocessing codebases to reduce noise, and organizing data for downstream tasks. We designed a system architecture and pipeline to support querying, response handling, and evaluation.