jam-actions-v1

Schema: jam-actions-v1/1.0.0 · Version: 1.1.0 · Records: 213 (154 train / 59 test, split by song) · Songs: 11 · Families: 9 · Licence: CC-BY-SA-3.0-DE · Source repo: mcp-tool-shop-org/ai-jam-sessions

The successor to jam-actions-v0. Where v0 asked whether a model could use the tools, v1 asks whether a small model can reason from what the tools return — and it exists in its current shape because, seven training runs in a row, the answer depended on what the assistant turn was asked to write.

What a record is

A multi-turn trace: a user question about a piece, the assistant's tool calls against the AI Jam Sessions MCP server (54 tools in the catalogue, 10 exercised here), the tool results, and a final assistant turn. Nine families:

family	n	the question	the final turn
acoustic	132	grade a recorded take: match, pitch_fail, timing_fail	cents 66.9: |66.9| − 50 = 16.9, against the gate; onset −9.8: |9.8| − 40 = −30.2, inside: pitch_fail
harmony	19	does a proposed reharmonisation clear the verifier	intended Bsus2, detected Csus2: different; chromatic 0/17 = 0.000 − 0.2 = −0.200, inside: rejected
measures	11	how many measures	the number
transpose	11	the key after transposing by an interval	the key
teaching_goals	11	the song's teaching goals	the list
chord	10	what chord is the left hand playing in measure N	the symbol
key_moments	8	the first key moment's measures	the range
ensemble	7	who entered first / which tone is wrong in a mixed take	the label
compare	4	do two pieces share a key (train only — the three held-out songs have three keys)	Eb major, F major: different: different_key

The acoustic, harmony and compare turns show their work: the measured quantity, the gate it is held against, the subtraction, the word the subtraction implies, then the label. The label is never produced by the arithmetic in the record; it is produced by the engine's predicate on the measured value, and the arithmetic is rendered beside it. A test asserts, for every record, that the printed numbers equal the tool result, the subtraction is exact, each word equals its predicate, and the label equals gold.

Nothing prompt-visible carries a threshold, a comparison word, or a class word. The gates (50 cents, 40 ms, a chromatic ratio of 0.2) appear only in assistant turns. The acoustic take paths are opaque hashes; the tool results are measurements at instrument resolution (0.1 cent, 0.1 ms).

The experiment contract this corpus was built under

1. Gold is constructible: every label is derived from a render or a computation the repo can redo.
2. Labels are verified against the engine that will grade them, not hand-written.
3. Split by the unit that leaks — the song — never by record.
4. Trivial baselines and the base model are reported on the same split as any adapter.
5. Thresholds live in the record, not in a grader's head.
6. Guard bands exceed estimator error (the acoustic tracker's onset p95 and pitch p95 are measured and recorded in tracker-error.ts).
7. A new corpus gets a new schema_version.

Plus three rules this arc added: gold must vary within every family, in train and in test; no answer in any prompt-visible field; and a tool-less baseline — a local 24B model given the user turn alone — must sit at the floor before any GPU is spent. It does: 10/40 held out, acoustic 5/17, at the three-way floor.

And one rule the arc earned last, at some cost: a song is in the corpus only if its arrangement's licence is verified from evidence — the download URL, the source's terms, and the MIDI file's own copyright and title events — recorded in the song's provenance block. The set is derived from those blocks by a filter, not typed by hand, and a test locks it at these eleven.

How it got this shape

Nine rebuilds, each a repair the previous run exposed. The numbers are held-out acoustic takes (the family where a small model had to compare a measurement to a gate); the base is Qwen2.5-3B-Instruct with a fair prompt, the adapter a rank-16 LoRA at the same recipe every time. Rows above the line were measured on a 27-song working corpus that a later provenance audit found unpublishable; the results were real measurements on real arrangements, and they are the reason the target looks the way it does, but that corpus is not this dataset.

corpus	what the target said	base	adapter	what the run showed
268, first build	bare label	—	—	five families with constant gold; withdrawn before training
268, repaired	bare label	27/69	33/69	p = 0.07; the acoustic prompt still named the class in the path
268, prompt de-leaked	bare label	9/27	10/27	the floor; the adapter never said match
268, comparison in words	cents 56.4 against a 50-cent gate …	10/27	20/27	above the floor — and every miss was a sharp take: it read the minus sign
349, onsets vary, both signs	comparison in words	20/54	47/54	flat 9/9, sharp 2/9: the sign was the whole gain
349, sign uninformative	comparison in words	20/54	38/54	pitch 2/18; a near-gate probe then showed the onset word followed the sign of the onset, never its size
349	digits of the subtraction	20/54	54/54	probe 70/72 (seed 13), 72/72 (seed 42); 7B 54/54 and 72/72
371, harmony/compare show theirs	digits	23/54	54/54	116/117 overall; probe 72/72; the base alone got harmony 13/14 and compare 6/6 once the tool returned the deciding quantities
146, eleven verified songs — 1.0.0	digits	3B 5/17 · 7B 7/17	3B 14/17, 12/17 (two seeds) · 7B 16/17, 17/17 (two seeds)	24-take probe: 3B 11/24 and 13/24 — at 48 training takes the 3B no longer locks the arithmetic or the vocabulary; 7B 24/24 at both seeds, every band 6/6, arithmetic exact on 21–22 of 24
213, the same eleven songs, four draws — 1.1.0, this release	digits	3B 13/36	3B 36/36, 36/36 (two seeds); on the 1.0.0 held-out takes 17/17, 17/17	probe 23/24 and 24/24; arithmetic exact on every held-out line at seed 42; no invented gate word. The limit was takes per song, not the number of songs
213, 1.1.0, Apache-2.0 bases	digits	Qwen3-4B 12/36 · 1.5B 12/36	Qwen3-4B-Instruct-2507 36/36, 36/36 (two seeds); 1.5B 34/36, 36/36	Qwen3-4B probe 24/24 and 23/24, arithmetic exact on every held-out line — the same numbers as the 3B on a base that permits commercial use, so it is the small adapter that ships. The 1.5B reaches the labels (24/24, 24/24) with sloppy digits (exact on 19 and 10 of 24). experiments/coverage-v1-sft/RESULTS-r52.md

On 1.0.0 the size of the training set was the limit for the smaller model: the 3B learned the format and copied the numbers but did not lock the comparison, at either seed; the 7B did, near the gate, on songs it never saw. 1.1.0 doubles the acoustic draws from the same eleven songs — 96 training takes instead of 48, every non-acoustic record identical — and the 3B locks it at two seeds. Three targets, one recipe, identical loss curves: a bare label trained a class prior, a worded comparison trained sign-reading, and the digits trained the comparison. The corpus never changed what the model was shown. It changed what the model was asked to write.

Full results, receipts and raw completions for every run are in the source repo under experiments/coverage-v1-sft/ (RESULTS*.md).

Files

file	what
records.jsonl	the 213 records, one per line
records/	the same records, one file each
splits.json	train/test ids; split by song (held out: solace, the-easy-winners, the-entertainer)
manifest.json, coverage.json	counts, tools, songs, shapes, floors — build artefacts, not claims
checksums.sha256	one entry per file, breadth-first, LF-pinned; verify with sha256sum -c
PROVENANCE-NOTE.md	which songs are excluded and why
LICENSE-DATASET.md, CITATION.cff	licence chain and citation

The acoustic takes are not shipped as audio. Each acoustic record's observation.acoustic block carries the phrase, the applied cents shift and delay, and the render's wav_sha256; the source repo re-renders every take from that block and re-derives every label from the render as part of its test suite (rebuild-equals-committed, numeric tolerance 1e-6). One caveat, measured: V8's transcendental functions are not correctly rounded across Node majors, so wav_sha256 can differ between Node 22 and 24 by last-place noise; the labels do not.

Evaluation companion

jam-actions-v1-probe — 24 acoustic takes on the three held-out songs measured within 10 ms or 5 cents of a gate, both signs, never trained on. It is the test that told a sign-reader from a comparator when the main split could not.

Source data and licence

Eleven public-domain compositions — Bach's Prelude in C (BWV 846), Beethoven's Für Elise, Mozart's Sonata K. 545 (first movement), and eight Scott Joplin rags — in arrangements from two sources: Bernd Krueger's piano-midi.de MIDI (three songs, CC-BY-SA-3.0-DE, the file's own copyright event naming him) and the Mutopia Project's LilyPond typesettings (eight songs, Public Domain, each piece's Mutopia page). Every record's provenance block carries the source URL, the terms it was read under, the licence, the arranger as the file names them, and the file's SHA-256.

The source library holds 108 songs; 97 are not here because their MIDI came from sites whose terms do not permit redistribution or whose licence could not be established. Three further songs are excluded by rule: clair-de-lune (the v0 fine-tune holdout), and Satie's Gymnopédie No. 1 and Debussy's Arabesque No. 1 (provenance unverified in the v0 audit). The exclusions are enforced by tests, and the audit that produced them is docs/findings/library-provenance-audit.md in the source repo.

The records, traces and annotations are released under CC-BY-SA-3.0-DE, the same licence as jam-actions-v0, so the two sets combine without a licence boundary; LICENSE-DATASET.md states the three layers. Code in the source repo is MIT.

Citation

See CITATION.cff. This release, jam-actions-v1 1.1.0, is Zenodo record 10.5281/zenodo.22679457, a version under the jam-actions concept DOI 10.5281/zenodo.20279918. 1.0.0 remains citable as 10.5281/zenodo.22675239. Cite the version DOI when you mean this dataset exactly; the concept DOI resolves to whatever the line's latest version is.

What changed in 1.1.0

• Four acoustic takes per (song, class) instead of two: 132 acoustic records (96 train / 36 test) instead of 65 (48 / 17). The 65 ids of 1.0.0 reappear with the same class and gold and freshly drawn values; every non-acoustic record is byte-identical to 1.0.0, in the same order.
• The coverage report's majority-shape floor (no family above 50% of records) is not met: the acoustic family is 62% of 1.1.0. The floor was written for a balanced corpus and is left where it is; this release records the miss in coverage.json rather than moving the gate, because the imbalance is the experiment (see the results table).
• CITATION.cff describes this dataset. The 1.0.0 file's abstract carried the previous working corpus's numbers (371 records, 27 pieces); the record itself, its card and its checksums were correct.