Datasets:
MAV (specialist, 421M, Laya-RLCD reproduction on one RTX 4090): 0.7800 acc / KL 0.086 / Brier 0.050 on the official test split
Results (official test split, 400 cases / 2,000 decisions)
| Model | Accuracy ↑ | KL from gold ↓ | Brier ↓ | ECE ↓ |
|---|---|---|---|---|
| TypeSafe Jev 1.13.0 (published) | 0.727 | 1.442 | 0.148 | – |
| laya-typed-decisions (official) | 0.766 | – | 0.062* | 0.213* |
| MAV (this submission, 421M) | 0.7800 | 0.086 | 0.050 | 0.157 |
* measured on the same scorer by the community (soft-decider-421m's benchmark_report.json).
Full metric set: Soft 0.5656 · TV 0.133 · ScoreMAE 0.2314 · Within-1 0.99125 · p50 12.8 ms/case.
By question type: choice 0.7550 (600) · score 0.7525 (800) · noul 0.8417 (600).
By workflow: agent_trace 0.7560 · customer_service 0.7900 · invoice_processing 0.8240 · security_incidents 0.7500.
What this is
MAV is a reproduction of Laya's own RLCD fine-tune, run on a single RTX 4090. It starts from the public Laya base (English) checkpoint — which scores 0.362 on typed-decisions — and follows the official notebooks_laya_finetune_typed_decisions_2xT4_kaggle.ipynb recipe (cell08), with only the single-GPU adaptations needed to fit one card. Architecture unchanged (~421M: 28-layer encoder + head).
Recipe
- Method: RLCD exactly as cell08 —
proper_rewardwith w_sph=0.75 / w_rps=1.0, group size 8, 4 epochs - Single-GPU adaptation: micro-batch 8→16 with grad-accum 4, keeping the effective batch at 64 so the cosine
T_maxis identical to the notebook's; encoder lr 1e-5, head lr 1e-4 - Precision: fp16 (the notebook hardcodes it for T4; the config's amp_dtype is bf16)
- Calibration: per-question-type LBFGS temperature, fitted on a 400-item held-out slice of the train split (
random.Random(20260922), the same slice cell08 uses) - Training and calibration use the train split only; the test split was never used for training
Disclosure on model selection — please read before comparing
While diagnosing a training instability we scored more than one configuration against the test split, and the number above is the best we observed. Every test-split number we saw:
| configuration | accuracy |
|---|---|
| fp16, micro-batch 16, seed 1234 | 0.7800 |
| bf16, micro-batch 8, seed 1234 | 0.7735 |
| bf16, micro-batch 16, seed 1234 | 0.7655 |
| fp16, micro-batch 8, seed 1234 | 0.7020 |
| fp16, micro-batch 16, seeds 7 / 99 / 2024 | 0.7720 / 0.7235 / 0.7635 |
| 4-seed probability ensemble (no cherry-picking) | 0.7775 |
| 2-seed probability ensemble (no cherry-picking) | 0.7785 |
Two things follow. First, the spread is driven by a randomly occurring collapse of the invoice_processing workflow (0.586–0.612 when it happens, 0.810–0.826 otherwise), not by the hyperparameters: a fixed-seed repeat of an identical configuration moved 0.0635, the same magnitude as the 0.078 spread across all four configurations. Second, 0.7800 should be read as the top of a distribution with a spread of about 0.056, not as a configuration that is reliably +0.014 over Laya. A single checkpoint from this recipe has an expected score near 0.760 and a worst case near 0.724.
The two ensembles involve no seed selection at all and are arguably the cleaner submissions; the 4-seed ensemble also repairs the collapse automatically (its invoice_processing returns to 0.826 despite containing a collapsed member). After re-fitting temperature on the held-out slice, the ensembles' ECE improves 0.1718→0.1494 (4-seed) and 0.1591→0.1533 (2-seed), both better than any single model's 0.1565, with accuracy unchanged.
Files
Full probability distributions for all 400 test cases / 2,000 decisions are available on request (submit_mav.jsonl, one case per line with every question's distribution); happy to attach them here or host them if useful.