Datasets:
Tasks:
Text Generation
Modalities:
Text
Formats:
json
Languages:
English
Size:
10K - 100K
ArXiv:
DOI:
License:
File size: 9,615 Bytes
d09f52e ad6fd64 d09f52e 3aa8392 d09f52e 83733b9 d09f52e 83733b9 d09f52e d1d5a6f d09f52e 83733b9 d09f52e 83733b9 d1d5a6f d09f52e 83733b9 d09f52e d1d5a6f 83733b9 d09f52e 83733b9 d09f52e 83733b9 d09f52e 83733b9 d1d5a6f d09f52e 83733b9 d09f52e 83733b9 d09f52e 83733b9 d09f52e 83733b9 0aa87cb 83733b9 0aa87cb 83733b9 0aa87cb 83733b9 0aa87cb 08bede3 3aa8392 08bede3 d09f52e 83733b9 d09f52e 83733b9 d09f52e 83733b9 d09f52e 08dc487 d09f52e 3aa8392 d09f52e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 | ---
pretty_name: "Rameau: Functional Harmony from Notation (Roman Numerals, Cadence, Key)"
license: cc-by-4.0
language:
- en
task_categories:
- text-generation
tags:
- music
- music-theory
- functional-harmony
- roman-numeral-analysis
- chord-progression
- cadence
- key-detection
- benchmark
- synthetic
- mir
- symbolic-music
size_categories:
- 10K<n<100K
annotations_creators:
- machine-generated
source_datasets:
- original
configs:
- config_name: symbol_to_rn
data_files:
- split: train
path: data/symbol_to_rn/train.jsonl
- split: validation
path: data/symbol_to_rn/validation.jsonl
- split: test
path: data/symbol_to_rn/test.jsonl
- config_name: notes_to_rn
default: true
data_files:
- split: train
path: data/notes_to_rn/train.jsonl
- split: validation
path: data/notes_to_rn/validation.jsonl
- split: test
path: data/notes_to_rn/test.jsonl
- config_name: pcset_to_rn
data_files:
- split: train
path: data/pcset_to_rn/train.jsonl
- split: validation
path: data/pcset_to_rn/validation.jsonl
- split: test
path: data/pcset_to_rn/test.jsonl
- config_name: key_id
data_files:
- split: train
path: data/key_id/train.jsonl
- split: validation
path: data/key_id/validation.jsonl
- split: test
path: data/key_id/test.jsonl
---
# Rameau: functional harmony from notation
A text-to-text dataset and benchmark for functional harmony: Roman-numeral
analysis, cadence classification, and key identification. A probabilistic
common-practice grammar generates the progressions; four task framings hide
the answer to increasing degrees. Chord-symbol lookup stops working after the
first one.
Named for Jean-Philippe Rameau, whose *Traité de l'harmonie* (1722) started
the discipline.
```
symbol_to_rn key: C major / progression: Dm7 G7 Cmaj7 -> ii7 V7 IM7 / cadence: PAC
notes_to_rn key: C major / notes: D4 F4 A4 C5 | ... -> ii7 V7 IM7 / cadence: PAC
pcset_to_rn key: C major / pitch classes: [2 5 9 0]|... -> ii7 V7 IM7 / cadence: PAC
key_id notes: D4 F4 A4 C5 | G3 B3 D4 F4 | ... -> C major
```
## Configs (tasks)
Load one with `load_dataset("4esv/rameau", "<config>")`. Default: `notes_to_rn`.
| config | task | rows |
|---|---|---|
| `symbol_to_rn` | key + chord symbols -> Roman numerals + cadence (easy: chord quality is given) | 5,715 |
| `notes_to_rn` | key + spelled notes -> Roman numerals + cadence (must read each chord) | 5,715 |
| `pcset_to_rn` | key + bass-first pitch-class lists -> Roman numerals + cadence (no spelling) | 5,715 |
| `key_id` | spelled notes, no key -> identify the key (only key-unambiguous phrases) | 4,795 |
| | **total** | **21,940** |
## Gold labels
Nothing is hand-annotated. The grammar (tonic -> predominant -> dominant ->
cadence, with sevenths, inversions, cadential 6-4s and secondary dominants)
generates each progression together with its intended analysis. Every chord is
then derived two independent ways with [`music21`](https://web.mit.edu/music21/):
the Roman-numeral figure through the roman engine, and the printed chord symbol
through the chord-symbol parser. An item is kept only if both agree on
pitch-class set and bass.
This release: 27,480 of 27,480 chords agree. See `VERIFY.md`.
Built from 845 progression shapes (key-independent), transposed across
keys. All content is synthetic; no third-party corpus is redistributed.
## Label convention
Roman numerals follow the feature decomposition of the
[DCML harmony standard](https://github.com/DCMLab/standards)
(`numeral` / `form` / `figbass` / `changes` / `relativeroot`).
We follow the notation and copy no DCML data. Major-seventh tonic is `IM7`;
secondary dominants use `/` (e.g. `V7/vi`).
Cadence codes: `PAC` perfect authentic, `IAC` imperfect authentic, `HC` half, `DC` deceptive, `PC` plagal.
## Fields
Common to every config: `input`, `target`, `key`, `mode`, `labels`, `cadence`,
`analysis` (per-chord DCML features), `source` (`grammar`/`curated`/`single`),
`category`, `shape_id`. Plus the input representation for the config: `chords`
(symbols), `notes` (spelled, bass-first), or `pitch_classes` (bass-first).
Accidentals are written the standard way (`Bb`, `F#`, `Cb`). music21 users:
its parsers want `-` for flats, so convert `b -> -` in note and root names
before calling `ChordSymbol` or `Pitch`.
## Splits
The atomic unit is a shape, a key-independent Roman-numeral sequence. A shape
hashes to exactly one split, so none of its transpositions or task framings
crosses splits. The test split doubles as the benchmark. Rows:
train 14,725 / validation 3,571 / test 3,644.
## Known limitations
- The distribution is synthetic. Grammar output, not repertoire; chord
statistics are not naturalistic.
- PAC vs IAC is decided by inversion, since there is no notated soprano.
Cadence rules are strict: an HC ends on a root-position V triad (a terminal
V7 is not labelled), and a DC requires a root-position dominant
(`V65 -> vi` does not count).
- key_id keeps only cadence-terminated progressions of three or more chords
whose notes contain scale degree 4 and the leading tone, so the key is
uniquely determined. Without the gate, gold keys are contestable: `I V7/V V`
in C is note-identical to `IV V7 I` in G, and the G reading is arguably
stronger. Such phrases are excluded.
- Harmony only: no voice leading, melody, or rhythm. No modal mixture,
Neapolitans, or augmented sixths yet.
## Evaluation
Gold is deterministic, so scoring is exact match. No LLM judge. The harness in
`eval/` is stdlib-only and works against any OpenAI-compatible endpoint:
```bash
python eval/run_model.py --config notes_to_rn --model <model> --out preds.jsonl
python eval/score.py preds.jsonl --config notes_to_rn --split test
```
Metrics for the RN configs: `exact` (numerals and cadence both correct),
`labels_exact`, `chord_acc`, `cadence_acc`. For `key_id`: `exact`, `tonic_acc`,
`mode_acc`. Prompts are versioned in `eval/prompts.py`; parsing rules are in
`eval/README.md`.
## Results
Full test split, zero-shot, temperature 0, prompt v1, run 2026-07-11 via
OpenRouter. Cells are exact match, with per-chord accuracy in parentheses.
Raw predictions and per-run metadata are in `results/`.
| model | symbol_to_rn | notes_to_rn | pcset_to_rn | key_id |
|---|---|---|---|---|
| gpt-oss-120b (reasoning low) | 0.256 (0.885) | 0.155 (0.778) | 0.146 (0.724) | 0.778 |
| Claude Sonnet 5 | 0.220 (0.863) | 0.066 (0.688) | 0.157 (0.732) | 0.823 |
| Qwen3-235B-A22B-Instruct | 0.193 (0.739) | 0.016 (0.380) | 0.002 (0.163) | 0.719 |
| DeepSeek-V3.2 | 0.151 (0.634) | 0.005 (0.305) | 0.001 (0.150) | 0.661 |
| Kimi-K2.5 | 0.142 (0.562) | 0.007 (0.399) | 0.009 (0.205) | 0.788 |
| Llama-3.3-70B | 0.037 (0.466) | 0.004 (0.307) | 0.000 (0.194) | 0.471 |
No model saturates the easiest config. Models that do not reason drop toward
zero once the chord symbols disappear; the two that do degrade more slowly.
Claude Sonnet 5 scores higher on pitch classes than on spelled notes, which
suggests spelling rather than harmony is its bottleneck. The table cost about
nine dollars in API credits.
### Reasoning on vs off
Fixed test subsets rerun with reasoning enabled; n in the table. Exact match
on `notes_to_rn`:
| model | n | off | on |
|---|---|---|---|
| Kimi-K2.5 | 150 | 0.000 | 0.740 |
| DeepSeek-V3.2 | 200 | 0.025 | 0.725 |
| Claude Sonnet 5 | 100 | 0.090 | 0.520 |
| gpt-oss-120b (low -> high) | 200 | 0.185 | 0.440 |
The pattern holds on every config; full numbers are in `results/reasoning/`.
With thinking enabled, per-chord accuracy reaches 0.89 to 0.98 on the hidden
configs, so the remaining exact-match gap is mostly cadence and figure errors.
Without thinking the benchmark measures pattern recall; with thinking it
measures multi-step computation. Neither saturates it.
## Reproduce
The full generation pipeline ships in this repo (`src/harmony_dataset/`):
```bash
uv sync && uv run pytest
uv run python -m harmony_dataset.export # regenerates data/, README, VERIFY.md
```
## Related work
- [MusicTheoryBench](https://huggingface.co/datasets/m-a-p/MusicTheoryBench)
(ChatMusician, 2024): 372 hand-written multiple-choice questions on broad
music knowledge. Rameau is generative and machine-verified.
- [Harmonic Reasoning in LLMs](https://arxiv.org/abs/2409.05521) (Kruspe, 2024):
synthetic interval, chord, and scale identification. No key context, so
identification rather than functional analysis.
- [Teaching LLMs Music Theory](https://arxiv.org/abs/2503.22853)
(Pond & Fujinaga, 2025): one RCM Level 6 exam in four encodings, with
prompting strategies. Rameau frames the same progressions in each
representation, so representation is the only variable.
- Score-based Roman-numeral analysis (Micchi et al., AugmentedNet, AnalysisGNN):
specialist models trained on NC-licensed annotated corpora. Rameau targets
text models and generates its own data, which is what keeps the license CC-BY.
## Licensing
CC-BY-4.0. Content is generated by this repository's pipeline from music theory;
the underlying facts are not copyrightable and no source corpus is redistributed.
## Citation
```
@misc{rameau,
title = {Rameau: Functional Harmony from Notation (Roman Numerals, Cadence, Key)},
author = {Stevens, Axel},
year = {2026},
doi = {10.57967/hf/9570},
url = {https://huggingface.co/datasets/4esv/rameau},
note = {Synthetic, music21-verified, DCML labels}
}
```
|