mcp-tool-shop/jam-actions-acoustic-v0

Dataset Card for jam-actions-acoustic-v0

Version: 1.1.0 Published at mcp-tool-shop/jam-actions-acoustic-v0. No DOI.

Summary

72 constructible gold records of grounded MCP tool use over monophonic audio analysis. Each record pairs a 4-note right-hand reduction of a public-domain library phrase with a seeded synthetic take and a gold verdict (match, pitch fail/warn, timing fail/pass, missed, extra, in-tune vibrato, or nothing-to-grade silence).

This is not a musical edition of the source pieces. A Bach measure contains far more than four notes. Reducing to four sequential right-hand onsets keeps the count honest: 9 perturbation kinds × 4 target notes = 36 records per phrase × 2 phrases = 72.

Source phrases

clair-de-lune is not in this corpus. It is the held-out test split of the published jam-actions-v0 fine-tune arc.

Compositions are public domain. Arrangement metadata is copied read-only from jam-actions-v0 (CC-BY-SA-3.0-DE, Bernd Krueger / piano-midi.de). Audio is original synthetic (fixtures-sine-v1), not those MIDI performances.

Split

Held out by phrase, not by record. Random record holdout would leak: the same phrase and kind at a different target note is nearly the same example.

• —Train: 36 records (Bach)
• —Test: 36 records (Für Elise)

Leaks: the nine-kind taxonomy, the tool sequence, and the gate numbers (they sit on every record). Does not leak: the held-out melody, its times, or which index was perturbed on that phrase.

Files

• —records.jsonl — one record per line, with split
• —records/ — the same records as individual JSON files
• —splits.json — phrase-locked split
• —manifest.json
• —VERSION
• —CITATION.cff
• —LICENSE-DATASET.md
• —checksums.sha256 — every other file, sorted breadth-first by path
• —This card

Every one of those is generated from the source repository, so the whole tree rebuilds from code.

Re-rendering the audio

No WAV files ship. Each take re-renders deterministically from observation.render.recipe, and wav_sha256 is the hash of the waveform it produces.

That hash is engine-dependent, and you should know before you check it. The renderer calls Math.pow and Math.sin once per sample, and ECMA-262 does not require either to be correctly rounded. V8's results changed between Node 22 and Node 24: of the 27,869 distinct Math.pow(2, x) arguments this corpus evaluates, 253 (0.91%) return a different double. Nearly all of that disappears under 16-bit quantisation, but 2 of the 72 records do not survive it — both the extra perturbation of Für Elise, whose motif sits on MIDI 63, the one pitch where the semitone ratio itself differs by a unit in the last place.

So:

• —Every other field of every record reproduces on any engine. Verify the download against checksums.sha256 and it matches everywhere.
• —`wav_sha256` matches on Node 22, the engine this corpus was generated on. On Node 24 expect those two records to differ. That is this, not a corrupt download.

Making the waveform bit-portable means replacing the transcendentals, which changes every hash and therefore every record. It would need a new schema_version, so it has not been done.

Version history

• —1.1.0 (2026-09-25). The 36 records built from schumann-traumerei were withdrawn. Earlier versions attributed that phrase to Bernd Krueger / piano-midi.de; the library's provenance audit of 2026-09-09 found the source file came from midiworld.com with no arrangement licence on record, so the attribution was wrong and the records could not stay under this dataset's terms. The Bach and Für Elise records are byte-identical to 1.0.2. The train split now holds one phrase (Bach) and the test split is unchanged (Für Elise).
• —1.0.2 and earlier. 108 records over three phrases. Superseded; do not redistribute the Träumerei records from those versions.

License

Traces and synthetic audio: CC-BY-SA-3.0-DE. Underlying compositions: public domain. See LICENSE-DATASET.md for the three layers and what each obliges.