Datasets:
Phocinae-Largha-150M-v1 — test scores (generalist) + reproduction
Tiny model. Big decisions.
We're releasing Phocinae-Largha-150M-v1 (model card), a 144.3M bilingual (zh/en) decision model for structured decisions — state / criteria / options in, one verdict out. mmBERT-small base, 256k vocab, runs on CPU. For local, open-source use — we don't run a service.
Scores (test, same protocol):
| Model | typed accuracy |
|---|---|
| Phocinae-Largha-150M-v1 — en, typed-decisions (400 cases / 2,000 decisions) | 0.797 |
| Phocinae-Largha-150M-v1 — zh, 400 cases (translated), no native Chinese training rows | 0.789 |
| Laya (published) | 0.766 |
| JEV (published) | 0.727 |
| meraGPT Decider1 (published) | 0.768 |
Mode: generalist.
What we think is the real story: shuffle the options, and the answer basically doesn't move — option-order flip rate is 0.0300 (reversed) / 0.0233 (random-mean) / 0.0433 (any), lower is better, bought by training every case in 5 random option orderings, not by data tricks. Brier ECE 0.1313. CPU p50 ≈1.5s per decision.
Our misses are public, too: JevBench public-231 aggregate 0.5108 (118/231) vs the 58.4% acceptance gate — not passed, disclosed as-is. Pre-registered gates G1–G4 all passed; details in the model card.
Reproduce: harness + frozen seed/row protocol · model card. Browser demo ships T+2 (local runtime: phocinae-server). Happy to take community head-to-heads on the same rows.
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 — an earlier version of this post said "generalist" and "no native Chinese training rows", and used 0.797 / 0.789 from the previous release. Updated wording and figures (v1.1, 2026-10-09) follow.
Tiny model. Big decisions.
We're releasing Phocinae-Largha-150M-v1 (model card), a 144.3M bilingual (zh/en) decision model for structured decisions — state / criteria / options in, one verdict out. mmBERT-small base, 256k vocab, runs on CPU. For local, open-source use — we don't run a service.
Scores (test, same protocol):
| Model | typed accuracy |
|---|---|
| Phocinae-Largha-150M-v1 — en, typed-decisions (400 cases / 2,000 decisions) | 0.906 § |
| Phocinae-Largha-150M-v1 — zh, 400 cases (machine-translated) | 0.848 § |
| Laya (published) | 0.766 |
| JEV (published) | 0.727 |
| meraGPT Decider1 (published) | 0.768 |
Mode: specialist (fitted on the typed-decisions train split; not a zero-shot generalist). The training mix includes machine-translated Chinese (≈2,400 rows) and native Chinese (≈1,400 rows) of its 10,000 items; the zh score is measured on machine-translated test cases — an in-mix (fitted) evaluation, not zero-shot Chinese transfer.
§ = self-scored on the official test split, per the card's reporting rules (maintainer-scored leaderboard rows are unmarked).
What we think is the real story: shuffle the options, and the answer basically doesn't move — option-order flip rate is 0.0217 (reversed; flip400 protocol; 0.0200 on flip150) / 0.0144 (random-mean) / 0.0283 (any), lower is better, bought by training on reordered-option mixtures. Calibration: shipped-column ECE 0.2519 (en; an optional calibration column bundled in calib/ reaches 0.0168; zh 0.1941 → 0.0152). CPU p50 ≈1.64 s per case.
Our misses are public, too: JevBench public-231 aggregate 0.5455 (126/231) vs the 58.4% acceptance gate — not passed, disclosed as-is. Pre-registered gates G1–G4 all passed; details in the model card.
Reproduce: harness + frozen seed/row protocol · model card · landing page: https://phocinae.github.io/Phocinae-Largha-150M-v1/. Happy to take community head-to-heads on the same rows.