Datasets:
Submission: Phocinae/Phocinae-Largha-150M-v1 — specialist (fitted on the train split); accuracy 0.906 (en) / 0.848 (zh)
Edited 2026-10-10: rewritten to the corrected v1.1 framing and figures; the correction comment below preserves the update record.
Phocinae-Largha-150M-v1 — 144.3M-param bilingual (en/zh) typed decision model on mmBERT-small (Apache-2.0), fine-tuned on this dataset's train split (with flip-augmented option reorderings). The training mix also includes machine-translated Chinese and native Chinese rows — please read the model as specialist (fitted on the train split), not zero-shot.
Results (all self-reported; details and eval code in BENCHMARKS.md):
- typed-decisions 0.906 (en) / 0.848 (zh, machine-translated cases) on the held-out test split (the train split was used in training) — the en score is above the 0.735 teacher self-agreement reference (the dataset card flags scores far above it as label-specific overfitting); JEV-27B reports 0.727 for comparison
- Mean model confidence (all decisions): 0.6541 en / 0.6546 zh
- Flip rate (mirror test): 0.0217 (release protocol, flip400)
- ECE: 0.2519 (shipped column; optional bundled calibration column in
calib/: 0.0168) — per-type temperature scaling - Latency: p50 21.0 ms on an RTX 5090 (fp16); about 1.64 s per case on CPU
- Confidence gate: at τ=0.6, 45.0% of decisions escalate to a larger model (the rest resolve locally, zero output tokens); LLM calls cut by 55.0% (79.6% at τ=0.5)
On JevBench we score 0.5455 (126/231), below the 58.4% acceptance gate — disclosed on the model card.
This is an official hf benchmark now, you can add the results to your model card and it should show up on the board, see - https://huggingface.co/docs/hub/en/eval-results
Correction — the original post said "zero-shot" and "trained from scratch"; corrected wording and updated figures (v1.1, 2026-10-09) follow.
Phocinae-Largha-150M-v1 — 144.3M-param bilingual (en/zh) typed decision model on mmBERT-small (Apache-2.0), fine-tuned on this dataset's train split (with flip-augmented option reorderings). The training mix also includes machine-translated Chinese and native Chinese rows — please read the model as specialist (fitted on the train split), not zero-shot.
Results (all self-reported; details and eval code in BENCHMARKS.md):
- typed-decisions 0.906 (en) / 0.848 (zh, machine-translated cases) on the held-out test split (the train split was used in training) — the en score is above the 0.735 teacher self-agreement reference (the dataset card flags scores far above it as label-specific overfitting); JEV-27B reports 0.727 for comparison
- Mean model confidence (all decisions): 0.6541 en / 0.6546 zh
- Flip rate (mirror test): 0.0217 (release protocol, flip400)
- ECE: 0.2519 (shipped column; optional bundled calibration column in
calib/: 0.0168) — per-type temperature scaling - Latency: p50 21.0 ms on an RTX 5090 (fp16); about 1.64 s per case on CPU
- Confidence gate: at τ=0.6, 45.0% of decisions escalate to a larger model (the rest resolve locally, zero output tokens); LLM calls cut by 55.0% (79.6% at τ=0.5)
On JevBench we score 0.5455 (126/231), below the 58.4% acceptance gate — disclosed on the model card.