jev-dllm
Reads Jev-style choice, yes/no and score probabilities from masked diffusion language models with shared Yes/No weights.
Overview
Jev-DLLM explores structured decisions with masked diffusion language models. It places a mask beside each candidate and reads Yes/No probabilities from shared weights to support dynamic choice, binary judgements and ordered scoring, without generating answer text. A released S1 checkpoint is based on Qwen3-0.6B, and documentation (in Chinese) covers temperature calibration, a five-model comparison with Kev, Dream zero-shot results and a Dream-S1 training run.
It includes a Core benchmark rebuilt from 14 fixed public sources into 17 sets and 6,744 decisions, with a data card noting exposure audits and unknown pre-training overlap. Evaluation runs on an NVIDIA GPU with BF16. The README is careful about provenance: the loader executes custom model code, requirements are not a full lock file, and validation reused an existing environment.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this candidate the right answer (shared Yes/No read)? | Yes · No |
| Score | Where does this item fall on an ordered scale? | — |
How it works
- Install the environment and download the S1 checkpoint
- Prepare the Core benchmark from public sources
- Place masks beside candidates and read Yes/No probabilities
- Evaluate choice, binary and scoring tasks and calibrate temperature
Strengths
- No text generation required
- Open checkpoint and benchmark preparation scripts
- Careful documentation of data exposure
Caveats
- Requires an NVIDIA GPU with BF16
- Loader runs custom remote code
- Research-stage results
Getting started
Create a Python 3.10 conda env, `pip install -r requirements.txt`, download SEU-ZZH/Shared-YesNo-Qwen3-0.6B-S1, then run benchmark/evaluate.py.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- PyTorch on CUDA GPU
- Has code
- yes
Use cases and integrations
How curated lists describe it
Shared Yes/No decision adaptation with diffusion language models.
hellogumbo/awesome-jev
Frequently asked questions
How does it make decisions?
It reads Yes/No probabilities at masks placed beside each candidate in a diffusion language model.
Is a model available?
Yes, an S1 checkpoint based on Qwen3-0.6B on Hugging Face.
What hardware do I need?
An NVIDIA GPU supporting BF16 with CUDA 12.1 drivers.
Related projects
- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython Open Jev-class decision models on Qwen3.5 with candidate-logit readout; the 4B model reports 80.1% on JevBench's 231 items.
ProjectresearchPythonOpen Jev replica on Gemma 4 E2B and Gemma 3 270M served on Modal, with eight demos and a TypeSafe eval head-to-head.
ProjectresearchPythonSmall experiment: reordering enum options changed a local Jev clone's answers on its old default, while hosted Jev's did not.
DemoresearchPython- JevK5★ 150
Apache-2.0 open-weight Jev alternative that answers yes/no, choice and score questions with per-option probabilities in one pass.
LibrarybetaPython Open System One decision model (LoRA on Qwen3.5-4B) that reads typed answers from logits in one pass, with its benchmark harness.
ModelbetaPython