Hurwitz/RNGen-data

RNGen data

This dataset provides the Key, Mixed, and RNA annotations and the processed score CSVs used for RNGen, a model for Roman numeral analysis. The source scores and harmonic analyses come from ten DCML corpora; the preprocessing builds on RNBERT.

Files

Directory	Contents
key/	train.json, valid.json, and test.json for local-key prediction
mixed/	The three JSON splits for the six-task intermediate training stage
rna/	The three JSON splits for Roman numeral analysis
processed_scores/	447 unaugmented score CSVs, their per-score JSON metadata, and ten corpus-level attrs.json files
preprocessing/	Sequence/segmentation and Octuple encoding settings

Use the released JSONs directly for training and evaluation with the RNGen code. The processed CSVs support inspection and data preparation. Keep the metadata JSONs with the CSVs.

The paper's score-level manifests select 355 training, 44 validation, and 47 test scores. Training augmentation covers 12 target keys per training score. The manifests define split membership rather than the directory contents alone. Phrase segmentation uses max_num_notes: 500 and adds one context bar on each side where available. Splits occur at bar boundaries; context notes and individual oversized bars can make the input exceed 500 notes. See processing details for the full augmentation description and modifications for the changes from the upstream corpora.

Download and use

Install the RNGen environment and activate it. From the root of the RNGen code checkout, download the data:

huggingface-cli download Hurwitz/RNGen-data --repo-type dataset \
  --local-dir datasets/releases/rngen

For fixed release revisions, use the code repository's download instructions. If authentication is required, first run huggingface-cli login with an account that has access. The corresponding model collection is RNGen-checkpoints:

huggingface-cli download Hurwitz/RNGen-checkpoints \
  --local-dir ckpts/releases

The standard code examples use datasets/releases/rngen/ and ckpts/releases/ inside the code checkout. A different data location works when passed explicitly as jsonl_dir during training or --json-path during evaluation. Baseline converters default to the standard datasets/releases/rngen/processed_scores/ directory.

Training and evaluation with the released JSONs do not read the source CSVs, their metadata sidecars, or the source paths recorded inside the JSONs. There is no need to regenerate inputs, transpose them again, or run the CSV preprocessing submodules for this route. Keep the CSV metadata files when using the separate data-preparation route.

Annotation format and split sizes

These are ordinary .json files containing RNGen-specific dictionaries, rather than JSON Lines or one flat list of records. Load them with Python's json module or RNGen's RNGenSFTDataset.

Stage	Train examples	Validation examples	Test examples	Training configuration	Task
Key	53,928	510	621	configs/rngen/key.yaml	key_sft
Mixed	323,568	3,060	3,726	configs/rngen/curriculum_stage2.yaml	rna_curriculum_stage2_sft
RNA	53,928	510	621	configs/rngen/rna.yaml	rna_sft

Each file includes a template entry with the task's prompt formatting. Examples use consecutive string keys, starting at "0":

• Key and RNA: examples contain events, spelling, note_count, instruction, output, and metadata. events is the prepared sequence of music event tokens. instruction identifies the target bars; the surrounding context is input only. Key outputs are text describing local-key regions. RNA outputs group harmony descriptions, chord tones, and chord-tone factors by key, and rna_info provides structured harmony information.
• Mixed: each example has seq_id, task_name, instruction, reference_input, output, and metadata. Its music input is stored once in the top-level seq_data dictionary and retrieved with seq_id. Templates for the six tasks are stored in template["tasks"].
• Metadata: source file paths and row identifiers record provenance; they are not paths that need to exist on the user's machine. last_target_bar records the final target bar. Mixed examples retain source metadata in metadata["source_metadata"].

Mixed supplies six examples per RNA segment, one for each task: harmony_span_prediction, chord_tone_inventory_prediction, tone_factor_pair_prediction, quality_prediction, inversion_figure_prediction, and ps_function_prediction. Thus its larger example counts do not represent additional independent scores.

Inspect a prepared RNA example without loading model weights:

import json
from pathlib import Path

path = Path("datasets/releases/rngen/rna/test.json")
with path.open() as f:
    annotations = json.load(f)
example_count = sum(key.isdecimal() for key in annotations)
print(example_count)  # 621
print(annotations["0"]["instruction"])
print(annotations["0"]["output"])

The released event sequences preserve the inputs associated with the checkpoints. Use them directly for reproducing results. The current converters support seven accidental tokens; regenerating annotations can change inputs containing triple accidentals. See the data-preparation guide before preparing new annotations.

Training and evaluation examples

Run from the code checkout with the RNGen environment active. To train Key from the supplied domain-fine-tuned RNBERT and Qwen:

python -m rngen.sft_ft --config configs/rngen/key.yaml \
  llm_model_name=Qwen/Qwen2.5-7B-Instruct \
  rnbert_checkpoint_path="$PWD/ckpts/releases/rnbert/unconditioned_checkpoint_best.pt" \
  jsonl_dir="$PWD/datasets/releases/rngen/key" \
  output_dir="$PWD/outputs/training/7B-Key" \
  report_to=none wandb_project=null

jsonl_dir is the stage directory despite the parameter's historical name; the entrypoint trains with train.json, validates with valid.json, and evaluates test.json after training. Use the matching Mixed and RNA configurations for later stages, setting finetuned_rngen_checkpoint to the best checkpoint from the previous stage and jsonl_dir to mixed/ or rna/. The full training guide shows the complete sequence and best-checkpoint selection. Use Qwen/Qwen2.5-14B-Instruct throughout the chain for 14B. The recipes use an effective batch size of 4 and require a suitable GPU platform.

To evaluate the released 7B-RNA checkpoint on the prepared RNA test split:

python -m rngen.sft_eval --config configs/rngen/evaluation.yaml \
  --model-path "$PWD/ckpts/releases/rngen/RNGen-7B-RNA" \
  --task rna_sft \
  --json-path "$PWD/datasets/releases/rngen/rna/test.json" \
  --output-dir "$PWD/outputs/evaluation/RNGen-7B-RNA"

Training accepts key=value overrides; evaluation accepts --flag value. Evaluation of all six released checkpoints, including the correct Key and Mixed task settings, is documented in the model card. Key and RNA evaluation print metrics and save prediction snapshots; Mixed evaluation reports the six task losses rather than full-RNA accuracy.

Processed scores and preprocessing settings

processed_scores/ contains unaugmented, note-level CSVs, not the segmented sequence CSVs consumed by the annotation builders. Its 447 scores include one score outside the 446 selected by the paper's split manifests. Select scores using data_splits/p_dcml/, which defines the 355/44/47 score split.

Keep each same-stem score JSON and each corpus's attrs.json beside its CSV. They record key, transposition, and corpus properties used by preprocessing. preprocessing/sequence_settings.json records sequence/segmentation settings; preprocessing/octuple_encoding_settings.json records encoding settings. Training transposition is performed before sequence encoding and segmentation; validation and test scores stay untransposed. The release contains the original processed scores and the prepared JSONs; augmented and intermediate CSVs are not included. See processing details.

License and source credits

The data and annotations are distributed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). The complete terms are in LICENSE.

Sharing and adaptation are permitted for noncommercial purposes subject to the license's attribution and ShareAlike conditions. Retain the original creator credits, source links, license notices, and descriptions of changes when redistributing this release or adaptations.

ATTRIBUTION.md lists all ten source corpora, their credited creators and archival identifiers, the applicable source licenses, the RNBERT preprocessing contribution, and the RNGen modifications. The MIT license of the RNGen source-code repository does not replace the data terms.

Citation

If you use this dataset in research, please cite the RNGen paper and the applicable upstream corpus and preprocessing reports in ATTRIBUTION.md. Machine-readable metadata is provided in CITATION.cff.

@inproceedings{gao2026rngen,
  author = {Gao, Hewei and Beyer, Tim and Wang, Xiaoxuan and Rohrmeier, Martin},
  title = {{RNGen}: A Large Language Model Approach to Roman Numeral Analysis Toward Explainable Harmonic Analysis},
  booktitle = {Proceedings of the 27th International Society for Music Information Retrieval Conference},
  year = {2026},
  url = {https://ismir2026program.ismir.net/poster_405.html}
}