mcp-tool-shop/jam-rollout-arc-evals

Rollout arc — raw generations

Every model generation behind the write-ups in mcp-tool-shop-org/ai-jam-sessions under experiments/rollout-arc/p4/.

Two things you can do with this.

Check our arithmetic. The repo has the readout scripts, the preregistrations and the intervals — but the generations they were computed from are ~51 MB and were never committed, so a clone got the conclusions and no way to recompute them. These are those files, unfiltered.

Or run the loop yourself. The trainer, the verifier bridge and the frozen progression set are all in the repo, and none of it is gated or fetched. Train your own adapter is the step-by-step: three commands, ~50 minutes on a 24 GB card, and the six traps that each cost us a run before we wrote them down.

The task

A 4B model proposes two-voice keyboard accompaniments for a fixed chord progression. A deterministic voice-leading checker grades each proposal against the full common-practice rulebook — parallel fifths and octaves, tendency-tone resolution, voice overlap, leaps, doubling. Reward is binary: admissible or not. No reward model, no judge, no human preference.

Every eval here is unconditioned (no scaffolding, nothing handed to the model), sampled G=64 on the same 75 held-out progressions, generation seed 7 unless noted.

Layout

Four configs, one per experiment family. Every split is the same 75 held-out items, 64 samples each — 1,500 rows in total.

config	split	what it is
base_probe (default)	chatml	Qwen3-4B-Base under a ChatML wrapper
	raw	Qwen3-4B-Base, no chat template
	fewshot	no template + one worked example, seed 7
	fewshot_s8	the same, seed 8 — the replication
local_gate	base	untrained Qwen3-4B-Instruct-2507
	c7l c8l c9l	plain GRPO, seeds 7 / 8 / 9
	pil7l pil8l pil9l	the --prefix-in-loss arm, seeds 7 / 8 / 9
	vs_base	Verbalized-Sampling control, G=8 × K=5
pod_b0cell	base	untrained, on the rented RTX 5090 with its own matched control
	c7 c8 c9	β=1e-4, seeds 7 / 8 / 9
	b07 b08 b09	β=0, seeds 7 / 8 / 9
prompts	heldout	the 75 prompts, with progression and the system/user text

prompts/fewshot-exemplar.txt is the single worked example the few-shot arm was given. It is a plain text file rather than a config — download it directly.

Loading

from datasets import load_dataset

# the headline arm: base checkpoint, no chat template, one worked example
d = load_dataset("mcp-tool-shop/jam-rollout-arc-evals", "base_probe", split="fewshot")
print(len(d), len(d[0]["completions"]))      # 75 items, 64 samples each

# the checkpoint it is measured against
load_dataset("mcp-tool-shop/jam-rollout-arc-evals", "local_gate", split="base")

Row format

{"itemId": "...", "songId": "...", "completions": ["...", ...], "lengths": [...],
 "distinct_exact": 64, "distinct_stripped": 64, "mean_entropy": null}

completions holds all 64 samples verbatim, including the ones that fail to parse. Nothing is filtered. The unparseable rate is itself a reported measurement in the base-probe arms (57.1% under ChatML, 25.9% un-enveloped, 2.0% with one worked example), so removing them here would delete a finding.

The prompts config is the one exception to that row shape: it carries itemId, songId, genre, chords, progression, and the system / user strings as they were sent.

Models

• Qwen/Qwen3-4B-Instruct-2507 — the instruction-tuned checkpoint the arc trained on
• Qwen/Qwen3-4B-Base — the pretrained sibling used in base_probe
• LoRA adapters are not published: they come from a cell whose own result was unresolved, and weights for a null are a download cost rather than a contribution. The how-to above produces your own in about fifty minutes.

The headline the data supports

Coverage — the fraction of items where at least one of 64 samples is admissible — is the quantity that matters when a verifier is in the loop, because best-of-n needs one hit:

	pass rate	items with ≥1 pass
instruct (trained on)	11.7%	39%
after GRPO training	12.3–12.9%	37% / 33%
base, no chat template	9.3%	93%
base + one worked example	16.2%	92%

Training moved coverage the wrong way. The instruct checkpoint emits fewer than two distinct openings per item out of sixteen admissible ones, so on 61% of items no amount of sampling reaches an admissible answer.

What this does not show: new capability (it is re-weighting what the base could already sample), that a prompt replaces training (those are different comparisons), or robustness to the choice of worked example (one exemplar was tried; the sampling seed was drawn twice, the exemplar was not). A fifty-line deterministic heuristic still scores 32/32 on the trained pool.

If you are going to train

Read the how-to first — it has the commands and the measured costs (22,946 MiB peak, ~50 min at 14.8 s/step). The short version of what bit us:

• --seed does not control LoRA initialisation. Two runs at one seed are independent draws, not replicates. Pin by artifact (--save-init-adapter / --init-adapter).
• One run tells you nothing. The same configuration produced +0.6 and +10.1 points.
• Start the bridge with --fixture and --require-pool. Otherwise a fresh clone builds a 14-song pool where a fetched machine builds a hundred — our first paid run was voided by exactly that, having served 14 rows to a trainer asking for 32.
• Unparseable output must score zero. The checker admits a realization with no sounding frames, so without a structure gate an empty completion earns full reward and the policy learns to emit nothing.
• Expect 46–76% of groups to produce no gradient. With a binary verifier, all-pass and all-fail groups have identically zero advantage.

Reproducing the published numbers

Clone the repo, drop these files into experiments/rollout-arc/p4/runs/ (and artifacts/ for pod_b0cell), then run the readouts — prior-shape.mts, base-passrate.mts, seed-stability.mts, why-unparseable.mts, pil-readout.mts. Every bootstrap seeds its RNG per computation, because a module-level generator consumed in call order once produced two different published intervals for the same quantity.

Licence

Apache-2.0, matching the repo and the base models. The prompts are chord-symbol charts and measure numbers already published in the source repository; the completions are model output.