JevAny
Open infrastructure for training, evaluating and deploying System 1 decision models, with calibrated LoRA checkpoints on a 27B backbone.
Overview
JevAny is open infrastructure for training and deploying System 1 decision models, covering data preparation, model adaptation and evaluation across language and multimodal backbones. One API takes a state, question and candidate options and returns a choice with probabilities, for uses such as routing support tickets, selecting tools or choosing a robot's next action. The current default release is a LoRA checkpoint on Qwen3.8-27B, with a Hugging Face collection, API docs, an interactive benchmark page and a case library of archived replays. Listings describe its training as calibration-aware reinforcement learning.
The README emphasises using a decision model inside LLM agent loops: delegate the selection bottleneck, not the task. Give it two to four valid branches with distinct outcomes, reuse one LLM plan across reversible choices, and scale autonomy only with paired reward and cost evidence. Demos show, for example, a WebShop task where Jev-style selection of color and size cut actions from 9 to 5, LLM calls from 9 to 4, tokens from 38,852 to 14,256 and time from 18.54 to 7.83 seconds, and a FrozenLake run with one LLM plan plus per-step choices. Documentation is in English and Chinese.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which of these 2-4 candidate actions should the agent take now? | — |
| Choice | Which team should handle this ticket / which tool should be selected?(inferred) | — |
How it works
- Prepare data and adapt a backbone with LoRA
- Evaluate calibration and accuracy
- Serve via one state/question/options API
- Use inside agent loops for bounded selections; escalate on novelty
Strengths
- End-to-end open training and serving stack
- Practical guidance on where to use decision models in agent loops
- Demos report large reductions in LLM calls and tokens
Caveats
- 27B backbone needs substantial hardware
- Demo replays come from an earlier checkpoint
At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- Training + serving stack
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Calibration-aware reinforcement learning for adaptive decision systems.
hellogumbo/awesome-jevInfrastructure for training, evaluating and deploying decision models across language and multimodal backbones, with calibrated LoRA checkpoints on a 27B backbone. English and Chinese.
MrJev/awesome-jev
Frequently asked questions
What is JevAny?
Open infrastructure for training, evaluating and deploying decision models, with released checkpoints.
How should it be used in agents?
For bounded selections among 2-4 valid branches, with the LLM handling planning and recovery.
What does the default model run on?
A LoRA on Qwen3.8-27B.
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 Joke decision model trained to answer wrongly with total confidence: 1.9% accuracy at 0.957 mean confidence, via a Jev-style API.
ProjectresearchPythonPre-registered field test of Jev, GLiNER2.5-Decide, Laya and other decision models against small LLMs on fresh September 2026 data.
ProjectresearchPython11 MB CPU-only model that answers choice, noul and score questions with fitted calibration and no LLM at inference.
ProjectresearchPythonTiny pre-registered experiment testing whether a confidence-scoring reward (RLCD-like) yields confidences safe to automate on at 0.9.
ProjectresearchPythonResearch code for an image-and-text System One model that reads option-letter logits from a multimodal LLM, with training and calibration.
ProjectresearchPython