Modelpublic

mcp-tool-shop/jam-actions-v1-qwen25-7b

sourceHugging Faceupdated 18d agoView on Hugging Face

0likes

Model Card

jam-actions-v1 adapters

Two LoRA adapters, one recipe at two seeds, trained on the released eleven-song jam-actions-v1 corpus. Every number below is on a held-out split by song or on the never-trained near-gate probe; the base model is reported on the same split; every seed that was run is reported — including the two 3B seeds that are not published, because on this corpus the 3B does not learn the comparison. Adapters trained during the arc on earlier, unpublishable versions of the corpus are not published either; their numbers are in the source repo's experiments/coverage-v1-sft/RESULTS*.md as the record of how this recipe was found.

Recipe

bf16 LoRA, r = 16, α = 32, dropout 0.1, on q/k/v/o/gate/up/down; lr 1.5e-4 cosine, 10 warmup steps; effective batch 8 (1 × 8 accumulation); three epochs; maxseqlen 16,384 (each example carries the full 54-tool catalogue, ~13.0–13.4k tokens); prompt-loss weight 0.1; chunked cross-entropy. Predictions greedy, max_new_tokens 128 (the shown-work line is 51–56 tokens). Receipts beside each adapter carry the seed, data SHA-256, hyperparameters, package versions and the loss curve.

Adapters

directory	base	seed	held-out acoustic	overall (40)	near-gate probe (24)	published
`7b-s13/`	Qwen2.5-7B-Instruct	13	16/17	37/40	24/24	yes
`7b-s42/`	Qwen2.5-7B-Instruct	42	17/17	38/40	24/24	yes
—	Qwen2.5-3B-Instruct	13	14/17	32/40	11/24	no
—	Qwen2.5-3B-Instruct	42	12/17	32/40	13/24	no
—	Qwen2.5-3B-Instruct, four-draw corpus, seed 13	13	17/17	38/40	23/24	not yet
—	Qwen2.5-3B-Instruct, four-draw corpus, seed 42	42	17/17	37/40	24/24	not yet

Bases on the same splits: 7B 7/17, 29/40, 12/24 (it says match to every probe take); 3B 5/17, 20/40, 6/24. The three-way floor is 5.7/17 and 8/24.

The two 7B seeds agree on every probe take (both 24/24, both exact on 21–22 of 24 subtractions, sharing the same rounding slip on one take) and differ by one held-out take; the shipped result replicates. The 3B rows are the reason the 3B trained on this corpus is not published: with ~45 acoustic training takes the 3B copies the numbers and keeps the format but does not lock the subtraction (seed 13: exact on 7 of 24 probe takes) or the gate vocabulary (seed 42 invents "within", "beyond", "behind"). On the earlier 349-record working corpus the same 3B recipe was run at two seeds (54/54 and 54/54 held out; 70/72 and 72/72 on the 72-take probe), which is the evidence that the recipe, not the seed, was doing the work there — and that the difference here is the size of the verified corpus. The last two rows say which size: a four-draw corpus of the same eleven songs (96 acoustic training takes instead of 48, every other record identical) brings the 3B to 17/17 held out and 23/24 and 24/24 near the gate at two seeds, with no invented gate word. The limit was the number of takes per song, not the number of songs. Those adapters are not published yet because the corpus they were trained on is not: it will be jam-actions-v1 1.1.0 when it is, and they follow it. Source repo experiments/coverage-v1-sft/RESULTS-r48.md.

What the adapters do and do not do

They write the comparison and then the label:

cents 66.9: |66.9| − 50 = 16.9, against the gate; onset −9.8: |9.8| − 40 = −30.2, inside: pitch_fail

Across every run of this recipe the line parses, the two numbers are copied from the tool result exactly, and the word follows the model's own subtraction. What the arc established is narrow and, we think, useful: the same 3B model at the same recipe learned nothing but a class prior from a bare label, learned to read a sign from a worded comparison, and learned the comparison from the digits — with identical loss curves in all three cases. If you fine-tune a small model to apply a rule, put the rule's arithmetic in the target and test near the boundary.

It was trained to compare against these gates (50 cents, 40 ms, chromatic ratio 0.2) in this tool vocabulary, on eleven songs. It is not a general grader.

Licence

The adapters are released under CC-BY-SA-3.0-DE, matching the dataset they were trained on. Use of a merged or adapted model is additionally governed by the base model's licence, and the two bases differ: Qwen2.5-7B-Instruct is Apache-2.0; Qwen2.5-3B-Instruct is under Alibaba's Qwen Research licence, which restricts commercial use. If commercial use matters, use this 7B, or the small path on Qwen3-4B-Instruct-2507 (Apache-2.0) at mcp-tool-shop/jam-actions-v1-qwen3-4b. Every adapter line we publish carries at least one permissively licensed base.

Reproduce

Source repo experiments/coverage-v1-sft/: scripts/train_v1_sft.py, scripts/predict_v1.py, scripts/score_v1.mjs, the run configs, and RESULTS-r40.md.