dpechenev/music-reasoning-benchmark

Procedural Music Reasoning Benchmark

This benchmark was generated with Danila-Pechenev/procedural-music-reasoning.

Benchmark version: v0.4.3. Generator version: 0.4.2.

It contains balanced examples from two implemented music reasoning task families:

• pitch_interval_reasoning
• chord_roman_reasoning

Configurations

Configuration	Examples per mode	Examples per split	Total examples
n16	16	256	768
n32	32	512	1536
n64 (default)	64	1024	3072
n128	128	2048	6144

The configurations are deterministic nested subsets in ascending size order: n16 is contained in n32 is contained in n64 is contained in n128. This makes results obtained at different benchmark sizes directly comparable.

Splits

Every configuration contains the same difficulty splits:

Split	Generator level
easy	0
moderate	3
hard	5

Columns

• id: stable row identifier.
• split: split name.
• level: generator difficulty/distribution level.
• difficulty: human-readable difficulty name.
• family: task family.
• mode: task mode within the family.
• prompt: model input.
• answer: canonical expected answer.
• answer_kind: answer-normalization family.
• cot: generator-produced reasoning trace.
• metadata: JSON string with symbolic generation metadata.

Evaluation Protocol

For benchmark evaluation, give the model the prompt only and compare its answer with answer using the task scorer. The cot field is provided for inspection, supervised training, and error analysis, but should not be included in the model prompt during benchmark evaluation.

Every benchmark prompt ends with Return only the requested answer, without explanation or additional text. This benchmark-only instruction requests the short answer expected by the scorer; it is not added to examples produced directly by the task generators.

Difficulty levels are distributional. A hard split may still contain some simple examples, but harder musical features are sampled more often or from a larger space.

Versioning

Hugging Face dataset releases should be tagged with the benchmark version. To load this exact release after upload, use:

from datasets import load_dataset

dataset = load_dataset(
    "dpechenev/music-reasoning-benchmark",
    "n64",
    revision="v0.4.3",
)

Replace n64 with any configuration listed above to select a different benchmark size.

Generation seed: 123.