cbjev
Self-hosted encoder model giving Jev-compatible choice, score and noul answers in one pass, pitched as a better-calibrated Laya successor.
Overview
cbjev answers typed questions (choice, score, noul) about any text or JSON state from one encoder pass, with every question in a call sharing one encoding of the state: 3 ms for one question and 11 ms for ten over a 500-token document on an RTX 4090, per the README. It is self-hosted, wire-compatible with Jev, and presented as a successor to Laya, using its own ModernBERT/mmBERT implementation that packs the state once followed by all questions.
The README compares cbjev and Laya on byte-identical cases: a mean accuracy of 0.741 versus 0.710 across 15 English suites, lower calibration error (ECE 0.117 vs 0.125) and far fewer answer flips when options are reordered (0.2% vs 7.8%), while Laya wins on some suites. Jev numbers are third-party published figures used only as indications because the author had no API access; on Banking77 with all labels in one question, Jev's published result is much higher. It is GPL-3.0 with weights on Hugging Face.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which label fits the text (e.g. AG News topic, intent)? | — |
| Yes / no | Yes/no over the state | yes · no |
| Score | Ordered rating over the state | — |
How it works
- Encode the state once with all questions
- Read typed answers for every question
- Serve over a Jev-compatible API
Strengths
- Millisecond latency per call
- Low option-order sensitivity
- Shared encoding for many questions
Caveats
- GPL-3.0
- Jev comparisons are indicative only
- Weaker than Jev on many-label questions per published figures
Getting started
Download weights from Hugging Face (0010101010-1/cbjev) and see the docs site.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- PyTorch server
- Has code
- yes
- License
- GPL-3.0
- Stars
- 4
Use cases and integrations
How curated lists describe it
Typed decisions (choice/score/noul) from one encoder pass - faster, better-calibrated successor to Laya, Jev wire compatible
wh000wh000/awesome-jev-live
Frequently asked questions
Is it compatible with Jev clients?
Yes, it is wire-compatible.
How fast is it?
About 3 ms for one question and 11 ms for ten on an RTX 4090, per the README.
Was it compared with Jev directly?
No, Jev figures are published third-party numbers.
Related projects
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.
ModelbetaPythonOpen Apache-2.0 System One decision models (4B, 9B, 27B on Qwen) with trainer, data builders, evals and every run record.
ServiceresearchPython