jam-actions-v1-probe

Schema: jam-actions-v1-probe/1.0.0 · Records: 24, all split: test · Evaluation only · Companion to: jam-actions-v1

Why it exists

An adapter trained on an earlier version of the corpus scored 47/54 on held-out acoustic takes. Its completions, which state the comparison before the label, showed that it wrote against a 50-cent gate whenever it saw a minus sign — and negative cents occurred in exactly one class of that corpus. The main split could not separate "compares |cents| to 50" from "reads the sign", because nothing in it measured near a gate. This set does.

What it contains

Three held-out songs (solace, the-easy-winners, the-entertainer) × four bands × both signs:

band	measured	gold
onset_in	|onset| ≈ 30 ms (±8), cents ≈ 20	match
onset_out	|onset| ≈ 50 ms (±8), cents ≈ 20	timing_fail
cents_in	|cents| ≈ 45 (±3), onset ≈ 2 ms	match
cents_out	|cents| ≈ 55 (±3), onset ≈ 2 ms	pitch_fail

Gold is derived from the two-sided predicates (|onset| > 40, |cents| ≥ 50) on the measured values, never from the intended band. Same record shape, same tools, same opaque take paths, same shown-work assistant turn as the main corpus. The takes were found by searching explicit applied delays and cents shifts; applied.json records what was applied and what it measured.

What it measured

On the 72-take version of this probe built over nine held-out songs of the earlier working corpus (the same bands, the same tolerances):

adapter (Qwen2.5, rank-16 LoRA, same recipe)	onset_in	onset_out	cents_in	cents_out	total
3B base, no adapter	6/18	1/18	8/18	5/18	20/72
3B, worded-comparison target	10/18	9/18	18/18	0/18	38/72
3B, arithmetic target (seed 13 / 42)	18/18	18/18	16–18/18	18/18	70–72/72
7B, arithmetic target	18/18	18/18	18/18	18/18	72/72

The worded-comparison adapter wrote inside for every negative onset (−56.2 ms included) and against for nearly every positive one (25.1 ms included). The arithmetic adapters write |56.2| − 40 = 16.2, against and |25.1| − 40 = −14.9, inside.

On this 24-take release, with adapters trained on the eleven-song corpus (106 train records, ~45 acoustic takes):

adapter	onset_in	onset_out	cents_in	cents_out	total
3B base	2/6	0/6	3/6	1/6	6/24
3B, arithmetic target, seed 13	4/6	0/6	3/6	4/6	11/24
3B, arithmetic target, seed 42	6/6	0/6	6/6	1/6	13/24
7B base	6/6	0/6	6/6	0/6	12/24 (says match to everything)
7B, arithmetic target	6/6	6/6	6/6	6/6	24/24

The 3B, which learned the comparison from ~110 takes on the earlier corpus, does not lock it from ~45: the subtraction is right on 7 of 24 (seed 13) and the gate vocabulary drifts (seed 42). The 7B learns it from the same 45: 24/24 parse, 24/24 copy the numbers, 22/24 subtractions exact, and the word follows the arithmetic every time.

Tool-less baseline (user turn only, mistral-small:24b) on the earlier 72: 18/72, below the three-way floor.

Rules

Never merge these records into a training set. Never split them. If you train on the companion corpus and report on this set, say which seed and report every seed.

Licence and citation

As jam-actions-v1: CC-BY-SA-3.0-DE for the records (see LICENSE-DATASET.md); the compositions are public domain and the three arrangements are the Mutopia Project's public-domain typesettings. This set is Zenodo record 10.5281/zenodo.22699571 under its own concept DOI, 10.5281/zenodo.22699570; cite either, not the jam-actions-v1 concept DOI — the probe is an evaluation companion, not a version of the training corpus. An earlier deposit of the same files, record 22675251, was minted as a version under the jam-actions-v1 concept by mistake and was deleted on 2026-09-10; its DOI resolves to a tombstone. Do not cite it.