openvons
openvons (open-Jev): finite-choice decision layer for text, images and Japanese voice, with a none option and execute/confirm/reject.
Overview
openvons, formerly open-Jev and named after the economist Jevons, makes LLMs, vision models and speech recognisers answer a finite set of options with a probability each instead of generating text. None of the above is always an option, and calibrated probabilities map to execute, confirm or reject. A shared core provides Noul, Choice and Score questions, confidence gating, calibration (temperature, isotonic, and with an explicit none option) and metrics such as ECE and Brier.
Modules cover language (a frozen 4B LLM with a trained head scoring 0.916 against 0.875 for a 27B zero-shot model, eight questions in 22.6 ms), vision (a frozen 407M encoder plus a 25k-parameter head beating a 27B model at a fraction of the VRAM) and Japanese voice commands (batch scoring of tens of thousands of proper nouns with kana-whisper, about 50 ms, 99-100% after calibration and full rejection of phone chatter). A related JevPick experiment speeds tool-call generation 3.2-4.8x by picking continuations from a menu. Figures are from the project's own docs.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which intent or tool does this request map to (or none)? | — |
| Choice | Which voice command or proper noun was spoken (or none)? | — |
| Choice | Gate on calibrated probability | execute · confirm · reject |
How it works
- Score finite candidates with a frozen model plus small head.
- Include a none option and calibrate.
- Map probability to execute, confirm or reject.
- Act or ask the user.
Strengths
- Explicit none option reduces forced wrong picks.
- Small trained heads beat much larger zero-shot models in its tests.
- Covers text, images and voice.
Caveats
- Results are from project docs.
- Voice module is Japanese-specific.
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- library
- Has code
- yes
Use cases and integrations
How curated lists describe it
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.
wh000wh000/awesome-jev-liveOpen decision layer for finite options across text, images, and Japanese voice commands.
cobanov/awesome-jevOpenvons (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.
hellogumbo/awesome-jev
Frequently asked questions
Why always include none of the above?
So the model can reject inputs that match no option, which is calibrated separately.
What inputs are supported?
Text, images and Japanese voice commands.
What is JevPick?
An experiment that speeds generation by picking likely continuations from a menu and verifying them.
Related projects
Research code for an image-and-text System One model that reads option-letter logits from a multimodal LLM, with training and calibration.
ProjectresearchPythonResearch fork of Laya that reads typed choice and yes/no decisions from an image plus optional text in one forward pass.
ProjectresearchPythonSelf-hostable calibrated decision model for images and text (Qwen3-VL-4B + LoRA) scoring 82.7 percent on public JevBench.
ProjectresearchPythonCalibration harness that reads yes/no, pick-one and rating answers from one forward pass of a frozen open vision-language model.
ModelresearchPythonResearch toolkit that encodes an image once and answers several typed questions from a shared distribution, plus OpenJev-0.6B text models.
ModelresearchPythonPaper introducing PixelJev, a native-image decision interface on small open multimodal models with few-shot adaptation.
Paperresearch