How do the text, vision and voice modules actually differ?

Text uses a frozen LLM plus trained head, vision uses a frozen 407M encoder plus small head, and voice scores candidates with kana-whisper plus calibration with a none option. Core provides Question, decide policy, temperature and isotonic calibration, and ECE, Brier, NLL and macro-F1 metrics.

Is JevPick a draft model or something else for generation?

JevPick builds a menu of likely continuations from tools, past outputs or the MTP head, picks from target hidden states and lets the target verify in one pass. Reported results are faster decode with byte-identical output including quantized transfer.

How does voice avoid firing the wrong station or camera name?

Voice transcribes to katakana readings, shortlists by edit distance under current UI state, scores candidates by teacher forcing in one batched pass, merges with the free hypothesis as none-of-the-above, then gates by risk. Registering a scope pre-trains names with TTS voices.

Can I point existing Jev code at a local openvons server?

The server exposes POST /v1/systemone with the same state plus questions shape as the TypeSafe SDK, so retargeting base URL answers locally. A Flutter app also runs recognition, scoring and decisions on-device after caching models.

What does the shared openvons.core layer provide?

The Noul, Choice and Score representation through Question, the decide policy with confidence gating, calibration (temperature, isotonic, and calibration with an explicit none option) and metrics including ECE, Brier, NLL and macro-F1.

How does the openvons voice pipeline pick a command?

kana-whisper transcribes freely into katakana, the UI state selects the valid command set, candidates are shortlisted by kana edit distance and scored by teacher forcing in one batched decoder pass, then combined with the free hypothesis as a none-of-the-above option and split into execute, confirm or reject.

Can the whole openvons pipeline run on-device?

Yes. The Flutter app runs log-mel, recognition, candidate scoring, calibration and the decision on the phone for voice, and image embedding plus choice matching for vision. The server only hands out models and the list of choices, so no audio or image leaves the device.

CommunityJev Practices & Patterns

Openvons

Decision layer answering finite text, vision and voice choices with none-of-the-above probabilities and execute, confirm or reject policy, plus a JevPick speculative decoder.

See highlights

Open github.com

Stars: 13
Language: Python
Published by: Community
Author: genai-craft
Last verified: Oct 2026

genai-craft/openvons

Organization repository on GitHub

View on GitHub

openvons (open-Jev): 有限選択肢に確率で答える判断層 — テキスト / 画像 / 日本語音声コマンド

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

Instead of asking an LLM, VLM or ASR model to generate text, openvons makes it answer a finite set of options with a probability for each. "None of the above" is always one of the options, and the calibrated probabilities are split into execute / confirm / reject. The name comes from Jevons (the marginal-utility economist). The project started as open-Jev; the package is openvons, and import jev still works as an alias. The same idea is implemented for three input types:

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Three modules sharing Noul, Choice and Score representation and calibration.
Voice pipeline with katakana hypothesis, batched scoring and risk-gated policy.
Vision heads for age, gender, orientation and baggage with calibrated levels.
Local systemone-compatible endpoint plus on-device Flutter app with cached models.

Watch out

Docs are partly Japanese and setup needs matching CUDA torch, vLLM and model hosts, so follow the language-specific design and evaluation pages before changing scopes. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-on-a-laptop

Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.