Datasets:

---

LocalLLaMA

/

typed-decisions

Like 87

Follow

LocalLLaMA 195

Mode: specialist. Fitted on the typed-decisions train split (1,080 states / 5,400 decisions; a further 120
states / 600 decisions of the train split held back for checkpoint selection). The test split was not used for
training or checkpoint selection of this model.

Model: mvbalaji/od1-typed-decisions — Qwen3.5-4B with
typed heads (choice / yes-no / ordinal CORAL) answering all questions of a case in one call; first trained as a
general decision model on openly licensed data with teacher labels, then fine-tuned on this benchmark's train split
with soft-gold targets (cross-entropy against each decision's gold distribution). Dataset revision
c76749ec58bd8c3d2ea706b31c333a9059c38f90.

Test results (400 cases / 2,000 decisions; full distributions: predictions_od1-typed-decisions.jsonl):

KL and Brier use the card's formulas (we reproduce its Uniform row exactly). Probabilities are renormalised over the
real answers (the model also has a NOT_ANSWERABLE outcome; mean mass removed 0.0001).

Calibration note. Against hard labels the model looks under-confident (answers given with 0.6–0.7 probability are
right 87% of the time on our held-back split); fitting a temperature to hard labels lowers ECE sharply but worsens KL
and Brier against the soft gold, so we report the untempered distributions.

Overlap disclosure. 94 of the 400 test states have ≥ 50% of their 13-grams somewhere in the train split
(template overlap); 26 have ≥ 50% inside a single train state (near-duplicate). Subset accuracies (supporting
information, not a replacement for the headline score): template-overlap 336/470 = 0.715, lower-overlap
1,257/1,530 = 0.822, near-duplicate 95/130 = 0.731.

Input-format disclosure. Four of the six yes/no question types carry definitions of true/false; this checkpoint
was trained and evaluated without them (instruction only). A corrected-input variant (definitions rendered, same
recipe) did not establish an improvement: 1,606 / 2,000 = 0.803 on test (paired difference +0.65 pt, 95% CI
[−0.45, +1.65], bootstrap over cases; its numerical gain was concentrated in the template-overlap cases, a pattern we
report without assigning a cause) and 902 vs 917 of 1,180 (−1.3 pt, CI [−2.6, 0.0]) on a separate set of four
workflows new to both models. That set was written and labelled by the same teacher model that labelled our general
training data and was not human-reviewed, so it measures agreement with a familiar teacher on new workflows, not
independently verified decision correctness. We therefore keep the released checkpoint.

Latency (our measurement). One idle H100 80GB, bf16, CUDA graphs, batch-1 requests, end-to-end answer() time
(tokenisation included), p50 over 40 repetitions: 10.3 ms for one question on a short state (≤ 64 tokens), 15.7 ms
on a 200–400-token state; 26.5 / 71.9 ms for five questions. On the same machine and workload,
laya-typed-decisions (native, all questions in one call) measured 16.1 / 17.1 ms (one question) and 18.2 / 18.9 ms
(five questions). These are p50 values; our p95 values mix in first-time input shapes (warm-up), so we make no p95
claim, and we do not report a serving-memory figure from this benchmark. (Per request, not per case of the card's
workload.)

Context (not matched reruns). For reference, author-reported specialist results in this discussion board /
model cards: soft-decider-421m 0.774 (author's harness; ~50 ms p50 on an unspecified GPU; its Brier/ECE definitions
are not stated, so not compared); Julia-1 0.7315 (author's report; training exposure to this train split unclear).
Our own reruns through a choice-wrapped interface (bare yes/no options): Jev 1.13 0.741, laya-typed-decisions 0.737
(0.769 in its native interface, which includes the yes/no definitions), Tev1-4B 0.690.

Caveats. The dataset card reports teacher self-agreement of 0.735 (a fresh teacher sample against gold built from
three other samples). A specialist fitted to the averaged gold can score above that; the benchmark measures agreement
with teacher-derived labels, not independently verified decision quality. This test split has also been used to
evaluate other models in our project (it was not used to train or select this checkpoint).

Artifacts: predictions · metrics + subsets · per-case overlap · evaluator code · README. Inference code and usage: the model card.