Datasets:
Tasks:
Text Classification
Formats:
parquet
Languages:
English
Size:
1K - 10K
Tags:
structured-decisions
calibration
probabilistic-classification
system-one
workflow-evaluation
Synthetic
License:
soft-decider-421m (specialist): 0.774 acc / 0.141 ECE on the official test split
#3
by winwinwinbb - opened
soft-decider-421m (specialist, fine-tuned on the train split)
RLCD fine-tune of convaiinnovations/laya (ModernBERT-large 421M) on this benchmark's own train split,
measured on the official test split (400 cases / 2,000 decisions) with a public harness
(the same harness reproduces laya-typed-decisions at 0.7664 against the published 0.766).
| model | acc | soft acc | Brier | ECE | score MAE | flip* | auto-decidable@5% | p50 ms |
|---|---|---|---|---|---|---|---|---|
| soft-decider-421m | 0.774 | 0.551 | 0.192 | 0.141 | 0.221 | 0.077 | 0.374 | 50 |
| laya-typed-decisions (same harness) | 0.766 | 0.500 | 0.213 | 0.214 | 0.243 | 0.065 | 0.378 | 50 |
| TypeSafe Jev 1.13.0 (published) | 0.727 | 0.580 | 0.148 | 0.144 | 0.391 | – | – | 710 |
* share of choice decisions whose argmax changes when the option order is reshuffled.
Mode: specialist — trained on train, scored on test, never zero-shot.
Temperatures (per-type + per-option-count buckets) are fitted on a 400-item slice held out of
training, which avoids the leak that leaves the upstream checkpoint over-confident (NandhaKishorM/laya#186).
Training recipe, eval scripts and the raw metrics are in the model repo (scripts/), single RTX 3090, ~17 min.