mcp-tool-shop/jam-actions-v1-qwen3-4b

jam-actions-v1 adapters, Qwen3-4B — the small path

Two LoRA adapters, one recipe at two seeds, on Qwen3-4B-Instruct-2507 (Apache-2.0), trained on jam-actions-v1 1.1.0 (the four-draw corpus of the eleven verified songs: 154 train / 59 test records, 96 acoustic training takes). This is the recommended small adapter: the same numbers as the 3B that found the result, on a base whose licence permits commercial use.

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. The 7B adapters (also Apache-2.0, trained on 1.0.0) are at mcp-tool-shop/jam-actions-v1-qwen25-7b; the tiny 1.5B option at mcp-tool-shop/jam-actions-v1-qwen25-1.5b.

Why this base

On 1.1.0 the recipe first locked the gate comparison at ~4B scale on Qwen2.5-3B-Instruct, whose base licence (Qwen Research) forbids commercial use. That result is published for the record at mcp-tool-shop/jam-actions-v1-qwen25-3b and is not the path to build on. Qwen3-4B-Instruct-2507 is Apache-2.0 on its own card, renders the corpus with the same span assertions on its own tokenizer, and reaches the same numbers. Source repo experiments/coverage-v1-sft/RESULTS-r52.md.

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, ~13.0–13.3k tokens); prompt-loss weight 0.1; chunked cross-entropy. Predictions greedy, max_new_tokens 128. Receipts beside each adapter carry the seed, data SHA-256, hyperparameters, package versions and the loss curve.

Adapters

Base on the same splits: 12/36, 6/24. Seed 42 is exact on 36/36 of its own held-out lines and 22/24 on the probe; its one probe miss is a subtraction slip that crossed the gate. Lines, seed 13: 36/36 of its own held-out takes parse, copy both numbers and subtract exactly; on the probe 24/24 parse, 22/24 exact (two slips on the cents term that touch no label), and the word follows the model's own subtraction and the true predicates on every line of every set.

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

They were trained to compare against these gates (50 cents, 40 ms, chromatic ratio 0.2) in this tool vocabulary, on eleven songs. They are not general graders.

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 here that is Apache-2.0: this is the commercial-safe small directory. Every adapter line we publish carries at least one such base; a non-commercial base is reported as an option, never as the only path.

Reproduce

Source repo experiments/coverage-v1-sft/: scripts/train_v1_sft.py, scripts/predict_v1.py, scripts/score_v1.mjs, lora-config-q3-4b.json, RESULTS-r52.md; the corpus is datasets/jam-actions-v1 at 1.1.0 (Zenodo 10.5281/zenodo.22679457).