Jebadiah
Open Apache-2.0 System One decision models (4B, 9B, 27B on Qwen) with trainer, data builders, evals and every run record.
Overview
Jebadiah (Jeb) is an open decision model that answers small typed questions about structured state and returns a calibrated probability over the allowed answers instead of text. It reads answers from the logits of single label tokens at one position, so there is no generation, parsing or refusal path. It supports choice, noul and score, aiming for answers that are always one of the allowed options, probabilities that mean something, stable answers under noise, and low enough cost to run on every event.
The repository holds everything needed to reproduce it: trainer and evaluator, data builders, evaluation harnesses, the configuration of every run and its result record. Models are trained on public data with AINode's trainer and served through AINode's /v1/decide and /v1/systemone routes. Weights at 27B, 9B and 4B on Qwen bases are published in bf16, GGUF and MLX under frontier-infra on Hugging Face, with a standalone server and playground and a llama.cpp script for GGUF builds.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which team should take this ticket? | — |
| Yes / no | Yes/no statement about the state | yes · no |
| Score | Place the state on an ordered rubric (e.g. five helpfulness levels) | — |
How it works
- Send state and typed questions to /v1/systemone
- Model reads label-token logits in one forward pass
- Return calibrated probabilities per option
- Caller acts, escalates or hands off by threshold
Strengths
- Fully reproducible: data builders, configs and run records included
- Apache-2.0 weights in several formats and sizes
- Jev-compatible API for drop-in local use
Caveats
- Larger sizes need substantial hardware
- Quality relative to Jev depends on task and size
Getting started
Download weights from the frontier-infra Hugging Face collection and run the standalone server, the llama.cpp script, or AINode.At a glance
- Type
- Service
- Status
- research
- Language
- Python
- Runtime
- local server, llama.cpp, MLX, AINode
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Open replica: Apache-2.0 decision models (27B, 9B, 4B on Qwen bases; bf16, GGUF and MLX) that answer `Choice`, `Noul` and `Score` questions with a probability for every option from one forward pass, and run anywhere: a standalone server with Jev's `/v1/systemone` wire and a playground, a llama.cpp script for the GGUF builds, or AINode (open-source local AI platform).
yibie/awesome-jevJebadiah, an open System One decision model: trainer, data builders, evals and every run record
logicrw/awesome-jev-projectsJebadiah, an open System One decision model: trainer, data builders, evals and…
wh000wh000/awesome-jev-liveApache-2.0 decision models (27B, 9B, 4B on Qwen bases; bf16, GGUF and MLX) that answer Choice, Noul and Score questions with a probability for every option from one forward pass. Runs anywhere: a standalone server with Jev's `/v1/systemone` wire and a playground, a llama.cpp script for the GGUF builds, or AINode (open-source local AI platform).
cobanov/awesome-jevJebadiah, an open System One decision model: trainer, data builders, evals and every run record.
hellogumbo/awesome-jev
Frequently asked questions
What is Jebadiah?
An open System One decision model that answers choice, noul and score questions with probabilities.
Is it compatible with Jev clients?
Yes, it is served through a /v1/systemone route.
Can I reproduce the training?
Yes. The repo includes the trainer, data builders, configs and results of every reported run.
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 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 Apache-2.0 local decision model (0.8B Qwen fine-tune on llama.cpp) behind the Jev request shape, calibrated and CPU-only.
ServicebetaPythonOpen 596M and 4B decision models that answer Choice, Score and Noul in one pass locally, with calibrated confidence for escalation.
ServicebetaPythonApache-2.0 118M multilingual decision model for Choice, Score and Noul on the System One API, with conformal abstain.
ModelbetaPython