Dataset Card for jam-actions-v0 (public subset)

Version: 0.5.0 Built: 2026-07-11 Source tag: jam-actions-v0-0.5.0-cut-2026-07-11 (record-content correction release — Bach BWV 846 errata 001 + 002; see RELEASE_NOTES.md for the full entry and evals/v0.5.0-release-gate-assessment.json + evals/v0.5.0-execution-verification.json for the gate verdicts at this state)

DOI: 10.5281/zenodo.20279918 — the concept DOI, which always resolves to the latest published version on Zenodo. This is the canonical citation handle. This release, v0.5.0, has version DOI 10.5281/zenodo.21313954 (published 2026-07-11, record at https://zenodo.org/records/21313954); the prior release v0.4.3 has version DOI 10.5281/zenodo.20279919 (published 2026-05-19).

📄 Fine-tuning evaluation — models trained on this dataset beat a prompted baseline at tool-grounded musical QA (+0.212, 29/36 wins, p < 0.0001). Read the preregistered confirmatory write-up: jam-actions-eval. Or explore it interactively — play the phrases and browse the results in your browser.

Dataset Summary

jam-actions-v0 is a corpus of multi-turn MCP (Model Context Protocol) tool-use traces grounded in real classical-piano MIDI. Each record pairs a short phrase window (typically 4 measures) with an annotated teaching target and a target trace — a turn-by-turn session in which an assistant uses the ai-jam-sessions MCP tools to read, analyze, and discuss the phrase. The dataset teaches LLMs to do grounded tool-use over symbolic music, not just text generation.

This is the public subset: 115 records across 8 compositions, all under CC-BY-SA-3.0 (DE jurisdiction). Two songs from the full source corpus (Satie Gymnopédie No. 1; Debussy Arabesque No. 1) are NOT included here — their provenance against piano-midi.de could not be verified during Slice 2.5 URL verification.

Dataset Structure

Top-level files:

• records.jsonl — one JSON object per line; the canonical training feed. Each line is a complete record with an additional split field ("train" or "test") so consumers can use the file without consulting splits.json.
• records/ — the same records as individual JSON files (sorted by id), useful for spot-inspection or downstream tools that prefer one-record-per-file.
• pianoroll/ — one SVG per record, matched by basename (<id>.svg corresponds to records/<id>.json).
• evals/ — eval artifacts produced by the source-repo eval harnesses and gates. The two canonical entry-point artifacts for this version are v0.5.0-release-gate-assessment.json (the RC-gate PASS verdict at the v0.5.0 cut) and v0.5.0-execution-verification.json (the standing execution gate: every unique frozen tool call replayed against the live MCP server, 0 failures). The slice-tagged artifacts (slice16–slice22) are the dataset's documented evolutionary history, all measured on pre-v0.5.0 record states — see KNOWN_LIMITATIONS.md §9 for the layered story. The Slice 21 fair-E3 baseline was measured on v0.4.3 records and ships in the v0.4.3 deposit (DOI 10.5281/zenodo.20279919), not here; the successor baseline measured on v0.5.0 records ships with a future version.
• splits.json — train/test split with held_out_song pinned. Locked: clair-de-lune is the canonical held-out test set; it is NEVER used for training.
• provenance-verification.json — per-song URL verification report from Slice 2.5 (filtered to the public songs).
• manifest.json — package-scope manifest with record_count, pair_count, songs_included, splits, etc.
• CITATION.cff — Citation File Format metadata.
• LICENSE-DATASET.md — layered-licensing explainer (public-domain compositions + CC-BY-SA-3.0-DE arrangements).
• VERSION — single-line package version.
• checksums.sha256 — SHA-256 sums of every other file in the package, sorted by path.
• package-inputs.json — packager-internal contract declaring which files are curated (preserved byte-for-byte) vs generated (rebuilt by scripts/package-jam-actions-public.ts on every run). Not consumed by downstream users; included for packager reproducibility.

Each record has these top-level fields: id, schema_version, provenance, scope, observation, annotation_target, target_trace, eval_metadata. See the source repo's src/dataset/schema.ts for the full Zod schema.

What's in a record

One line of records.jsonl is one phrase-window record. Abbreviated below (Clair de Lune, the held-out test song; … marks omitted prose/notes):

{
  "id": "clair-de-lune:m001-004:piano:mcp-session:v1",
  "split": "test",
  "scope": { "song_id": "clair-de-lune", "phrase_window": "measures 1-4", "key": "Db major",
             "tempo_bpm": 100, "time_signature": "9/8",
             "musical_phrase_label": "opening atmosphere — introductory phrase" },
  "provenance": { "composer": "Claude Debussy", "arrangement_creator": "Bernd Krueger",
                  "arrangement_license": "CC-BY-SA 3.0", "record_verdict": "public",
                  "training_use_permitted": true /* + source URL, PD status, verifier, verified_at */ },
  "observation": {
    "midi_sidecar": { "ticks_per_beat": 480, "timed_events": [
      { "t_seconds": 0.3, "dur_seconds": 2.4, "note": 65, "name": "F4",
        "velocity": 36, "hand": "right", "measure": 1, "beat": 0.5 } /* … one per note … */ ] },
    "tokens_remi": "…", "tokens_abc": "…", "piano_roll_svg_inline": "<svg…>"
  },
  "annotation_target": {
    "structure": "Opening atmospheric introduction — 9/8 ripple, floating long melody notes",
    "key_moments": ["m1 opening chord with sustained melody", "…"],
    "teaching_goals": ["9/8 triplet pulse must feel like three big beats, not nine", "…"],
    "style_tips": ["…"], "teaching_notes": [{ "measure": 1, "note": "…", "technique": ["…"] }]
  },
  "target_trace": {
    "task_family": "analyze-and-play-phrase",
    "objective": "Read mm. 1–4 of Clair de Lune, view the piano roll, analyze the phrase, …",
    "session": [
      { "turn": 1, "role": "user",      "content": "Show me measures 1–4 of Clair de Lune and describe the opening atmosphere." },
      { "turn": 2, "role": "assistant", "content": "Let me view the piano roll for mm. 1–4.",
        "tool_calls": [{ "tool": "view_piano_roll", "arguments": { "songId": "clair-de-lune", "startMeasure": 1, "endMeasure": 4 } }] },
      { "turn": 3, "role": "tool", "tool": "view_piano_roll",
        "content": { "svg_returned": true, "measures": 4, "rh_notes": 32, "lh_notes": 0 } }
      /* … assistant → tool → assistant, ending in a grounded analysis … */
    ]
  },
  "eval_metadata": { "split": "test", "leakage_check": "…", "eval_eligibility": "…" }
}

• scope locates the phrase — song, measures, key, tempo, meter, and its role in the piece.
• observation is the phrase itself in three synchronized views: a MIDI sidecar (timed_events: note, timing, velocity, hand, measure/beat), REMI + ABC token strings, and a rendered piano-roll SVG.
• annotation_target is the human teaching layer — structure, key moments, goals, style tips, per-measure technique notes.
• target_trace is the supervised signal: a multi-turn MCP session where the assistant calls the inspector tools on this phrase and answers from their results. This is what teaches grounded tool-use, not prose recall — and it's the surface the fine-tuned model learns.
• provenance carries the full license + verification chain; eval_metadata carries the leakage-checked split.

Load it

The dataset viewer renders every record (nested JSON), and the datasets library loads it directly:

from datasets import load_dataset

ds = load_dataset("mcp-tool-shop/jam-actions-v0", split="train")   # all 115 records; split tag is per-row
rec = ds[0]
print(rec["id"], "·", rec["scope"]["musical_phrase_label"])
print("notes:", len(rec["observation"]["midi_sidecar"]["timed_events"]),
      "· trace turns:", len(rec["target_trace"]["session"]))

# clair-de-lune is the locked held-out test song — never used for training:
held_out = ds.filter(lambda r: r["split"] == "test")

(records.jsonl carries both train and test records, each tagged with its split; splits.json pins clair-de-lune as the held-out song.)

Reproducibility

Earned by Slice 23.5 (audit-driven cleanup): a fresh contributor cloning this repo on any platform (Windows native, macOS, Linux, WSL) should be able to verify the package's integrity and reproduce the canonical Slice 22 RC-gate PASS verdict without operator handholding.

Package version pinned by this reproducibility section: 0.5.0 (built 2026-07-11; Bach BWV 846 correction release, errata 001 + 002).

Canonical tags:

• jam-actions-v0-0.5.0-cut-2026-07-11 — this version's cut (record-content corrections + the standing execution gate).
• jam-actions-v0-rc-gate-revised-2026-05-19 — Slice 22 RC-gate revised state (axes 2 + 6 revised).
• jam-actions-v0-feature-marketed-2026-05-19 — the v0.4.3 published state (Zenodo DOI 10.5281/zenodo.20279919).

Verification steps:

# 1. Clone and check out this version's tag.
git clone https://github.com/mcp-tool-shop-org/ai-jam-sessions.git
cd ai-jam-sessions
git checkout jam-actions-v0-0.5.0-cut-2026-07-11

# 2. Install deps (pnpm 10+ recommended; npm/yarn also work via the lockfile).
pnpm install

# 3a. Verify package checksums (274 files, ~2 seconds).
#     Exit 0 on success; exit 1 with `[bad line] / [hash mismatch] / [missing on disk]`
#     lines on failure. Windows-safe since Slice 23.5: `.gitattributes` pins LF
#     for *.sha256 files, and the verifier is CRLF-tolerant as defense in depth.
pnpm exec tsx scripts/verify-public-package-checksums.ts

# 3b. Re-run the standing execution gate: every unique frozen tool call in the
#     package replays against the live MCP server (0 failures required).
#     Requires a machine with an audio device — play_song connects the real
#     audio engine, so headless CI boxes report devicenotavailable instead.
pnpm build
pnpm exec tsx scripts/verify-public-package-execution.ts

# 3c. Reproduce the v0.5.0 RC-gate PASS verdict. The gate's input is the sealed
#     16-record E3 baseline, which was measured on v0.4.3 records and ships in
#     the v0.4.3 deposit rather than in this package (see RELEASE_NOTES.md).
#     Restore its sealed bytes from git history, then run the pure validator:
git show jam-actions-v0-feature-marketed-2026-05-19:datasets/jam-actions-v0-public/evals/slice21-fair-e3-baseline-results.json > /tmp/slice21-baseline.json
pnpm exec tsx scripts/check-release-gate.ts /tmp/slice21-baseline.json

Expected outputs:

The checksum verifier ends with [ok] All checksums verify, every file accounted for. and exits 0.

The execution gate prints VERDICT: PASS (230 unique calls, 0 failures) and exits 0 — matching evals/v0.5.0-execution-verification.json.

The release-gate CLI prints a per-axis summary with all 6 blocking axes PASS and aggregate RC gate PASS, exit code 0 — the structural verdict recorded in evals/v0.5.0-release-gate-assessment.json. (Honest scope: that baseline predates this version's prose corrections; the successor baseline measured on v0.5.0 records ships with a future version.)

Regression-check (the Slice 19 baseline should still FAIL under the revised gate):

pnpm exec tsx scripts/check-release-gate.ts \
  datasets/jam-actions-v0-public/evals/slice19-fair-e3-baseline-results.json

Exit code 1, with axes 1/2/6 reported as FAIL (the Schumann m045-048 record's pre-Slice-21 annotation was the catastrophic-stratum failure that Slice 21 repaired). This artifact is referenced as evals/slice22-release-gate-slice19-regression-check.json for the canonical regression-check verdict.

No model runs are required for reproducibility. The release-gate is a pure validator over sealed eval artifacts, and the execution gate replays deterministic tool calls. To re-run the model itself, see the source repo's scripts/eval-jam-actions-annotation-grounding.ts (requires Ollama + the qwen2.5:7b model locally).

Source Data

MIDI arrangements are by Bernd Krueger, published at piano-midi.de under CC-BY-SA 3.0 (DE jurisdiction). The underlying compositions are all in the public domain in both the US and EU (composer death + 70 years elapsed; latest of the 8 composers, Debussy, d. 1918).

Songs included (alphabetical):

• bach-prelude-c-major-bwv846
• chopin-nocturne-op9-no2
• chopin-prelude-e-minor
• clair-de-lune
• fur-elise
• mozart-k545-mvt1
• pathetique-mvt2
• schumann-traumerei

Licensing

This dataset is layered:

1. Compositions — public domain (US: pre-1929 publication; EU: composer death + 70 years elapsed).
2. Arrangements (MIDI sequences) — Bernd Krueger, piano-midi.de, CC-BY-SA 3.0 (DE) — https://creativecommons.org/licenses/by-sa/3.0/de/
3. Derivative records (this dataset) — CC-BY-SA-3.0-DE — share-alike inherited from the upstream arrangements.

HuggingFace's enumerated license slugs do not include the -de jurisdiction; the dataset card YAML uses cc-by-sa-3.0 and the DE jurisdiction is documented here in the body and in LICENSE-DATASET.md.

Attribution requirements when using this dataset:

• Cite Bernd Krueger and piano-midi.de when using the MIDI bytes or sequences.
• Cite this dataset (see CITATION.cff) when using the records, traces, or derived tokenizations.

Held-out Test Set

clair-de-lune (12 records) is the canonical held-out test set. It is never to be used for training. The remaining 103 records form the train split. The held-out choice is stratified by composer + style era: Debussy's Impressionist (1905) voicing is distinct from every training-set composer's idiom, so leakage from train to test is structurally low.

Split discipline is preserved across the packaging: every pair (prompt + continuation_target) is in the same split, and clair-de-lune was held out from the start of v0.

Provenance

Slice 2.5 of the source repo verified the provenance URL for each song against piano-midi.de's per-composer page. Of the 10 candidate songs, 8 passed verification and were promoted to public; the other 2 were demoted to internal (excluded from this public subset). The per-song report — including verification timestamps, attempted URLs, response sizes, and failure reasons (for the demoted songs) — is shipped alongside this dataset as provenance-verification.json (filtered to the 8 public songs).

Each record's provenance block carries its own verdict_reason, verifier, verified_at, and arrangement_evidence_url byte-for-byte from the source corpus. None of these fields are re-derived during packaging.

Citation

See CITATION.cff for machine-readable metadata. BibTeX equivalent:

@dataset{jam_actions_v0_public_2026,
  author       = {mcp-tool-shop-org},
  title        = {jam-actions-v0 — AI Jam Sessions tool-use traces (public subset)},
  version      = {0.5.0},
  year         = {2026},
  license      = {CC-BY-SA-3.0-DE},
  url          = {https://github.com/mcp-tool-shop-org/ai-jam-sessions}
}

Limitations

• Satie Gymnopédie No. 1 and Debussy Arabesque No. 1 are NOT in this public subset. Slice 2.5 could not confirm their provenance against piano-midi.de — the Satie composer page returned HTTP 418 (the upstream does not currently carry Satie), and the Debussy composer page was reachable but did not reference Arabesque No. 1. Both songs remain in the source repo with record_verdict: "internal" and are excluded from this distribution.
• No vocal records. The source repo's manifest.json declares vocal_synth_engine as a declared_dependency_surface for future record types (vocal_phrase, sing_along_trace, phoneme_alignment, vocal_render_score), but v0 ships only instrument records. The public-subset manifest.json reflects this by listing only ai_jam_sessions under instrument_surfaces.
• Piano only. All 8 songs are solo-piano arrangements. Other instruments are out of scope for v0.
• English-only annotations. Teaching-note text is English-only.

Maintainer

mcp-tool-shop-org — please open an issue at https://github.com/mcp-tool-shop-org/ai-jam-sessions for questions, corrections, or contributions.