Organization repository on GitHub
Turn any open model into a classifier/jev endpoint
- Forks
- 70
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Repos & SDKs
Simple Jev Project
A local server and Python package that runs open models with Hugging Face Transformers or PyTorch and returns Choice, Score, and Noul answers read from next-token logits instead of generated JSON.
- Stars
- 608
- Language
- Python
- License
- APACHE-2.0
- Reading time
- 3 min
- Published by
- Community
- Author
- featherless-ai
- Last verified
- Sep 2026
About this project
Use compatible open models from huggingface, for structured classification and scoring, without training a separate classifier head. Explore the demos, playground, and documentation at simple-jev.featherless.ai. Send shared context and a set of questions. Simple Jev reads the model's next-token logits for each question and builds a JSON response containing choices, rubric scores, or truth/support judgments. The model does not generate a JSON completion: the server constructs the response from the scores.
Quoted from the project's README. Read the full README.
Works forJev for Scoring
Used inJev for Software & DevTools
Highlights
- The server builds JSON responses from next-token logits, so nothing is generated and usage.output_tokens is always zero.
- A public demo API needs no login or key, with a 2k-token context limit and 2 requests per second; Featherless paid plans raise the limits.
- Backends cover a Hugging Face server, a native Laya Typed Decisions encoder with a 1,024-token-per-question default, and RFDT fine-tuning scripts.
- usage.input_tokens counts unique token prefixes and shares the common context across the questions in a request.
- Ships a playground and documentation at simple-jev.featherless.ai plus a shared common/ module for validation, prompt planning, and scoring.
Quickstart
bash
git clone https://github.com/featherless-ai/simple-jev.git
cd simple-jev
python3 -m venv .venv && source .venv/bin/activate
python -m pip install -e './hf-server'
python hf-server/hf_server.py --model Qwen/Qwen3.5-0.8B --device cpu --dtype float32Watch out
Apache-2.0. Needs Python 3.12+ and downloaded weights; the demo API is limited to 2k tokens and 2 RPS, and calibration varies by model.
Questions about this entry
Short answers drawn from the linked project or page.
No. It reads the next-token logits of an ordinary open model for each question and builds the JSON response from those scores, so the model is never asked to generate JSON. That does not make the probabilities good - calibration varies by model and the choices are yours to set.
It is a demo: a 2,000-token context limit and 2 requests per second with no key at all. Featherless paid plans raise both limits. It already serves a Gemma classifier model id, and GET /v1/models lists what you can pass to POST /v1/classifier.
No, over-limit requests are rejected with 422. The limits are startup flags: --max-request-branches defaults to 100 questions, --max-model-len to 16384 tokens per rendered question branch, and --max-choice-options to 255. Raising input length does not extend the checkpoint's native context or memory.
You can skip it. With no --classifier-prompt-policy flag the loader matches backbone type, depth, widths, attention dimensions, vocabulary and expert configuration, then auto-selects a development-selected format. Unrecognized configurations fall back to baseline with a startup warning to run prompt tuning first.
Use Python 3.12 or newer. Clone the repo, create and activate a venv, pip install -e './hf-server', then run python hf-server/hf_server.py with a model and device flags. The server listens on http://127.0.0.1:8000, with /health and interactive docs at /docs.
--max-request-branches caps questions per HTTP request at 100 by default, one scoring branch each, with 256 the schema maximum. --max-model-len defaults to 16384 tokens per rendered question branch. --max-choice-options defaults to 255, configurable 2 to 255, while Score stays limited to 50 levels. Over-limit requests get 422.
The loader matches the model's language-backbone architecture and size configuration, checking backbone type, depth, widths, attention dimensions, vocabulary and expert configuration, to development-selected formats. Unknown configurations use baseline with a prominent startup warning to run prompt tuning first, and an explicit flag always wins.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
Open multilingual System 1 decision models with published checkpoints for choice, score and noul questions, plus a router that dispatches each request to the right checkpoint in one forward pass.
Jev Repos & SDKs#community#python#open-models
Community
NandhaKishorM/layaPython
NandhaKishorM/layaPythonAn Apache-licensed System One decision server that reads Choice, Score, and Noul probabilities from DiffusionGemma 26B-A4B through vLLM on NVIDIA or MLX on Apple silicon, so TypeSafe SDKs work unchanged.
Jev Repos & SDKs#community#python#open-models
Community
razorback16/openjevPython
razorback16/openjevPythonA reproduction of Jev that turns any Qwen checkpoint into a decision model serving the same /v1/systemone schema (Choice, Score, Noul), with no training and no generated answer text.
Jev Repos & SDKs#community#python#open-models
Community
zhengxuyu/litjevPython
zhengxuyu/litjevPythonAn async Python client for Jev with typed Noul, Choice, and Score helpers: questions batch into one request, and callers get probabilities, choices, and weighted scores instead of text to parse.
Jev Repos & SDKs#community#python#sdk
Community
AboveColin/jevclientPython
AboveColin/jevclientPythonAn open, non-autoregressive System One model with published weights and a TypeSafe-compatible /v1/systemone API; the 395M-parameter OptionMarker checkpoint scores discrete and ordinal questions locally.
Jev Repos & SDKs#community#python#open-models
Community
wfzyx/vonPython
wfzyx/vonPythonTurns local text and vision models into Jev-style decision models by reading probabilities from prefill logits alone, with SGLang, Transformers, and Apple Silicon MLX backends.
Jev Repos & SDKs#community#python#open-models
Community
Yinsongxu/LLM2JevPython
Yinsongxu/LLM2JevPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.