What is sokudan in one sentence?

A Japanese System One decision model: send text plus typed questions and get typed answers with probabilities, without generating a single token. It is trained and evaluated on Japanese and presents itself as a Jev-style model rather than a chat model.

Which question types does it support?

choice picks an option from a list and returns probabilities plus confidence; score places the state on an ordered scale with a legend; noul returns the probability that a statement is true and has no confidence field. The schema is free per request, with no retraining.

How do I install and run it?

Python 3.11, then pip install sokudan. Load the model with sokudan.load('GeneLab/sokudan-ja-310m'), pass a state string and a questions dict, and read result['answers']['department']['choice']. The README's worked example routes a duplicate-billing complaint to the billing department.

What hardware does it need?

At 314.6M parameters it runs on a CPU. On Apple silicon with macOS 14 or later, pip install sokudan brings MLX with no torch; elsewhere it brings torch and picks cuda, then mps, then cpu. An M1 Max handles a 5-option choice in about 13 ms with MLX, 1.4 to 1.7 times faster than torch on MPS.

Any tips on the input format?

Pass the state as a string. The README warns that a dict is rendered as key: value lines, which is not the input the model was trained on, and the output changes. An explicit backend or device does not fall back automatically, and each self-check failure during auto selection is a warning followed by the next candidate.

How does it compare with laya-multilingual?

The README states the model is trained and evaluated on Japanese and beats laya-multilingual on every metric on Japanese business messages, and the project includes bench_ja, bench_en, and a reproduction of Laya's position-bias issue. The digest does not include the tables, so read the repository for the numbers.

Who built it, and what license applies?

hiroki-abe-58 publishes it under Apache-2.0, with release v0.3.0, model weights at GeneLab/sokudan-ja-310m, a PyPI package, and bilingual docs including README_ja. The MLX speed figure is from one idle M1 Max, so treat it as a single-machine measurement.

Jev directory

CommunityJev Models & Reimplementations

Sokudan (即断)

A Japanese System One decision model built on modernbert-ja-310m: send text and typed questions and get typed answers with probabilities from one forward pass, no text generation, at 314.6M parameters that also run on a CPU.

See highlights Open github.com

Stars: 4
Language: Python
License: APACHE-2.0
Reading time: 3 min
Published by: Community
Author: hiroki-abe-58
Last verified: Oct 2026

hiroki-abe-58/sokudan

User repository on GitHub

View on GitHub

Japanese System One decision model (Jev-style): typed answers and probabilities in one forward pass, no text generation. Includes bench_ja/bench_en and a Laya position-bias repro.

Forks: 1
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

Tagscommunity catalog-import cobanov japanese decision-model modernbert

Works forJev for Classification Jev for Scoring

Used inJev for Software & DevTools

Languages: Python
Platforms: macOS API
Capabilities: Open sourceRuns locally

Highlights

Answers are read off a decision head's logits: no text to parse, no JSON to validate, and no way to answer outside the options you sent.
Backbone is sbintuitions/modernbert-ja-310m; weights are published as GeneLab/sokudan-ja-310m at 314.6M parameters.
One forward pass per question with no decoding loop; on an M1 Max a 5-option choice takes about 13 ms with MLX.
Platform split: MLX with no torch on Apple silicon with macOS 14+, torch elsewhere across cuda, then mps, then cpu.
Question types are choice, score, and noul (bool is an alias); the schema is free per request with no retraining.

Quickstart

python

import sokudan

agent = sokudan.load("GeneLab/sokudan-ja-310m")
questions = {"department": {"type": "choice", "instructions": "この問い合わせはどの部署が担当すべきか",
                            "criteria": {"請求": "支払い・返金", "技術": "不具合・障害", "営業": "料金・新規契約", "その他": "上記以外"}}}
result = agent.predict("先月の請求で同じ金額が二回引き落とされています。至急ご確認ください。", questions)
print(result["answers"]["department"]["choice"])

Watch out

Grounded in the repository README and GitHub metadata reviewed on 2026-10-15; the claim that sokudan beats laya-multilingual on Japanese business messages follows the README's bench_ja/bench_en tables, the MLX speed comes from one idle M1 Max, and the backend-detail section was truncated in the digest.

Questions about this entry

Short answers drawn from the linked project or page.

From the Jev guides

Original write-ups on what this entry is for.

Beyond text and boxes: multimodal decisions you can run yourself

Oct 16, 2026

Jev-Omni's merged classifier scores text, image, audio, and video options; VideoJev and Laya Vision show what open multimodal decision models accept, what they return, and which numbers are the authors' own.

Jev Models & Reimplementations#multimodal#open-models#vision

Read article

Jev 1.14: Images, Box, and the Danger Zone

Oct 15, 2026

Jev Day shipped a routine model bump and a deliberate widening of the API: images on both models, a box question type on 1.14 only, and 262k context, all behind danger-zone headers.

Jev Models & Reimplementations#release#vision#danger-zone

Read article

The Open Decision-Model Roundup

Oct 14, 2026

Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.

Jev Models & Reimplementations#open-models#training#local

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai

Sokudan (即断)

Highlights

Quickstart

Questions about this entry

What is sokudan in one sentence?

Which question types does it support?

How do I install and run it?

What hardware does it need?

Any tips on the input format?

How does it compare with laya-multilingual?

Who built it, and what license applies?

More Jev resources like this

From the Jev guides

From the Jev AI model community