Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

Phocinae Largha 150M

Phocinae Largha: 144.3M bilingual English/Chinese decision model on mmBERT-small with calibrated Choice, Score and Noul in one pass.

ModelresearchApache-2.0
View on GitHub ↗

Overview

Phocinae-Largha-150M-v1 (the spotted seal) is a 144.3M-parameter decision model built on mmBERT-small, English-first with Chinese evaluated on machine-translated cases. It takes a state plus typed questions (noul, choice, score with 2-10 levels) and returns calibrated answers with confidence in one forward pass, trained for robustness to option reordering. Weights are Apache-2.0 on Hugging Face, GitHub and ModelScope.

The README reports 0.906 on English typed-decisions as a specialist fitted on that dataset's train split (noting the dataset card warns scores far above 0.735 may indicate overfitting), 0.848 on machine-translated Chinese, option-order flip rates around 2%, 21.0 ms p50 on an RTX 5090 and about 1.64 s per five-question case on a single CPU thread. It discloses 0.5455 on public JevBench, below its own gate. An escalation gate at 0.6 keeps low-risk decisions local and cuts LLM calls by 55% while raising kept-subset accuracy. Shipped ECE improves sharply with the bundled calibration column.

What the decision model decides

TypeQuestionAnswers
Yes / noYes/no question about the stateyes · no
ChoicePick one option—
ScoreScore on 2-10 levels—

How it works

  1. Load the model locally.
  2. Send state and typed questions.
  3. Read calibrated answers and confidence.
  4. Escalate to an LLM below the 0.6 gate.

Strengths

  • Small, bilingual, runs on CPU or GPU.
  • Robust to option reordering.
  • Honest disclosure of weak JevBench result.

Caveats

  • Typed-decisions score is a specialist fit, possibly overfit.
  • Chinese results on machine-translated cases.

Getting started

Download Phocinae/Phocinae-Largha-150M-v1 from Hugging Face; serve it with phocinae-server.

At a glance

Type
Model
Status
research
Language
Python
Runtime
PyTorch
Has code
yes
License
Apache-2.0

Use cases and integrations

How curated lists describe it

  • Open model: a 144.3M-parameter bilingual (English/Chinese) typed decision model on mmBERT-small that returns calibrated `Choice`, `Score` and `Noul` probabilities in one forward pass so callers gate or escalate on the numbers; runs locally, 21.0 ms p50 per decision on an RTX 5090, with Apache-2.0 weights.

    yibie/awesome-jev

  • Phocinae's bilingual (English/Chinese) typed decision model built on mmBERT-small (144.3M params, Apache-2.0): one forward pass yields calibrated probabilities for Choice, Score and Noul, trained for option-order invariance, with a confidence gate that keeps low-risk decisions local. typed-decisions 0.906 (en, specialist: fitted on this dataset's train split) / 0.848 (zh, machine-translated cases); p50 21.0 ms on an RTX 5090 (fp16) and about 1.64 s per case on CPU; at τ=0.6, 55.0% of decisions resolve locally with 0 output tokens.

    AnotiaWang/awesome-decision-models

  • Independent bilingual (English/Chinese) typed-decision model of 144.3M parameters on mmBERT-small, Apache-2.0, that returns calibrated Choice, Score and Noul answers in one forward pass and runs locally at 21.0 ms p50 per decision on an RTX 5090.

    cobanov/awesome-jev

Frequently asked questions

How big is it?

144.3M parameters on mmBERT-small.

How does it do on JevBench?

0.5455 on the public 231 items, which the README discloses as below its gate.

Which languages?

English first, with Chinese evaluated on machine-translated cases.

Related projects

  • cbjev★ 4

    Self-hosted encoder model giving Jev-compatible choice, score and noul answers in one pass, pitched as a better-calibrated Laya successor.

    ProjectresearchPython
  • ej

    11 MB CPU-only model that answers choice, noul and score questions with fitted calibration and no LLM at inference.

    ProjectresearchPython
  • kev★ 9,079

    Family of small Jev-like decision models you can train and run yourself, with a server compatible with TypeSafe's SDK.

    ServicebetaPython
  • Gutsy

    Apache-2.0 local decision model (0.8B Qwen fine-tune on llama.cpp) behind the Jev request shape, calibrated and CPU-only.

    ServicebetaPython
  • TinyJev

    Open 596M and 4B decision models that answer Choice, Score and Noul in one pass locally, with calibrated confidence for escalation.

    ServicebetaPython
  • Verdict

    Apache-2.0 118M multilingual decision model for Choice, Score and Noul on the System One API, with conformal abstain.

    ModelbetaPython