MetaScore-SFT-100

Overview

MetaScore-SFT-100 is a fixed, 100-piece benchmark for symbolic piano continuation, derived from MetaScore and developed to evaluate DAMN.

Each example pairs a 4-bar piano prompt (with an optional pickup) with a 12-bar reference continuation, following the MIREX 2026 Symbolic Music Generation format. This is an independently curated development benchmark, not the official MIREX test set.

Dataset

The benchmark contains 100 complete examples across 15 genre labels, with fixed prompts, references, and source assets.

Resource	Contents
test.jsonl	100 entries with CID, genre, source, and prompt/reference note sequences
challenge_test/	Fixed prompt/reference JSON, companion MIDIs, and window manifest
midi/, mscz/, mp3/	Complete-piece MIDI, MuseScore scores, and existing source previews
label/	Genre annotations
manifest.csv, summary.json	Source selection metadata and statistics
attribution.json	Per-piece attribution and source license records
checksums.sha256	File integrity checksums

Notes use integer (start, pitch, duration) attributes on a sixteenth-note grid. Steps 0-15 are reserved for an optional pickup, prompt starts are in 0-79, and reference starts are in 80-271. Durations are positive and notes end no later than step 272.

Source permissions apply per piece; see LICENSE.md. The existing source-license conflict is documented in rights_conflicts.json.

Construction

We selected piano scores from a genre-stratified sample of MetaScore using these criteria:

• A single piano part, 4/4 meter, at least 16 measures, and no lyrics.
• The upstream best-unique-arrangement filter.
• A user rating of at least 4.5 from at least three ratings.

Automatic filtering retained 106 candidates; manual listening removed six clear quality-control failures, leaving 100 pieces. Genre counts are not balanced.

Deterministic extraction selects the first complete, protocol-compatible 16-bar window: 99 pieces use the opening window, and one uses a later window. Fixed-window MIDIs use 480 ticks per quarter note, 120 BPM, and a fixed piano program. Complete-piece source assets retain their original contents; their MP3s are not renderings of the extracted windows.

See RECONSTRUCTION.md for reproduction instructions and DETAILS.md for full selection, timing, and provenance details.

Usage

from datasets import load_dataset

benchmark = load_dataset("naivecynics/MetaScore-SFT-100", split="test")
assert len(benchmark) == 100

example = benchmark[0]
prompt = {"prompt": example["prompt"]}
reference = {"generation": example["reference"]}

For reproducible experiments, pin the dataset revision and keep generation settings fixed across systems.

DAMN evaluation protocol: generate eight continuations per prompt, compute each continuation's CLaMP3 cosine similarity to its reference, average across the eight candidates, then average across the 100 prompts. The 95% confidence intervals use bootstrap resampling over prompts. This is mean-of-eight, not best-of-eight. Reference similarity is a development proxy, not a substitute for listening evaluation.

Citation

Please cite the original MetaScore work and this benchmark:

@inproceedings{xu2025metascore,
  title={Generating Symbolic Music from Natural Language Prompts Using an LLM-Enhanced Dataset},
  author={Xu, Weihan and McAuley, Julian and Berg-Kirkpatrick, Taylor and Dubnov, Shlomo and Dong, Hao-Wen},
  booktitle={Proceedings of the International Society for Music Information Retrieval Conference},
  year={2025}
}

@misc{huang2026metascoresft100,
  title={MetaScore-SFT-100: A Fixed Piano Continuation Development Benchmark},
  author={Huang, Langchu and Xu, Yang},
  year={2026},
  howpublished={Hugging Face dataset},
  url={https://huggingface.co/datasets/naivecynics/MetaScore-SFT-100}
}

Related resources: DAMN code · Model weights · MIREX 2026 results · Listening demo

Maintainer: Langchu Huang (langchuhuang@gmail.com).