ej
11 MB CPU-only model that answers choice, noul and score questions with fitted calibration and no LLM at inference.
Overview
ej is an 11 MB model that makes typed decisions on CPU with no LLM and no tokens generated. You give it a state (free text or JSON as text) and a set of typed questions, and it returns one probability distribution per question in a single pass, supporting choice with any option texts chosen at run time, noul yes/no statements and score over ordered levels.
The model file contains the tokenizer and encoder config, so no base model is downloaded, and inference is deterministic for a fixed thread count. Calibration is fitted and then measured per suite against the error a perfectly calibrated model would show. Weights use a pickle-free format checked section by section before decoding. It supports few-shot adaptation from a handful of labelled records and training your own model. Code is GPL-3.0 and weights CC BY-SA 4.0; a technical report is forthcoming.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which team should handle this ticket? | billing · tech |
How it works
- pip install ejai and load the model
- Pass state and typed questions
- Read per-question probabilities
- Optionally adapt with a few labelled records
Strengths
- Tiny, CPU-only and deterministic
- Pickle-free, verified weights
- Few-shot adaptation
- Measured calibration
Caveats
- Copyleft licences (GPL-3.0 code, CC BY-SA weights)
- Small model limits complex reasoning
- Technical report not yet published
Getting started
pip install ejai, then ej.load('5ak3t/ej', revision='v0.0.1').predict([...]).At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- CPU library
- Has code
- yes
- License
- GPL-3.0
Use cases and integrations
How curated lists describe it
An 11 MB model that makes typed decisions on CPU (choice, yes/no, score) with…
wh000wh000/awesome-jev-live
Frequently asked questions
How big is ej?
About 11.4 MB in one file including tokenizer and config.
Does it need a GPU?
No, it is CPU-only.
Can I adapt it to my data?
Yes, with few-shot adaptation from a handful of labelled records.
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 - 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.
ModelbetaPythonOpen Apache-2.0 System One decision models (4B, 9B, 27B on Qwen) with trainer, data builders, evals and every run record.
ServiceresearchPython