BMdataset: A Musicologically Curated LilyPond Dataset

TL;DR

This paper introduces BMdataset, a curated LilyPond music dataset, and LilyBERT, a specialized encoder, demonstrating that small, expert datasets can outperform large noisy corpora in music classification tasks.

Contribution

The creation of a curated LilyPond dataset and a domain-specific encoder, showing improved music understanding with small, high-quality data.

Findings

Fine-tuning on BMdataset outperforms pre-training on larger corpora.

Combining pre-training and fine-tuning yields highest accuracy (84.3%).

Small curated datasets can be more effective than large noisy ones for music classification.

Abstract

Symbolic music research has relied almost exclusively on MIDI-based datasets; text-based engraving formats such as LilyPond remain unexplored for music understanding. We present BMdataset, a musicologically curated dataset of 393 LilyPond scores (2,646 movements) transcribed by experts directly from original Baroque manuscripts, with metadata covering composer, musical form, instrumentation, and sectional attributes. Building on this resource, we introduce LilyBERT (weights can be found at https://huggingface.co/csc-unipd/lilybert), a CodeBERT-based encoder adapted to symbolic music through vocabulary extension with 115 LilyPond-specific tokens and masked language model pre-training. Linear probing on the out-of-domain Mutopia corpus shows that, despite its modest size (~90M tokens), fine-tuning on BMdataset alone outperforms continuous pre-training on the full PDMX corpus (~15B tokens) for both composer and style classification, demonstrating that small, expertly curated datasets can be more effective than large, noisy corpora for music understanding. Combining broad pre-training with domain-specific fine-tuning yields the best results overall (84.3% composer accuracy), confirming that the two data regimes are complementary. We release the dataset, tokenizer, and model to establish a baseline for representation learning on LilyPond.