Jev directory
CommunityJev Models & Reimplementations
UltraBERT Jev 380M (sqrt-softplus)
Fine-tuned UltraBERT-380M decision model on Hugging Face that reads a state, question, and option list and picks one option in a single forward pass. Scores options with a square-root softplus of yes minus no, normalized over options; 73.2% dev and 71.9% test accuracy over 13 datasets.
- Reading time
- 2 min
- Published by
- Community
- Added
- Nov 2026
Highlights
- Reads a state, a question, and a list of options, then picks one option in a single forward pass without generating text.
- Option k is scored from masked-token logits as square-root softplus of " yes" minus " no", divided by the sum over options.
- Trained for one epoch at batch 32 and learning rate 1.5e-5 on the 13-dataset Reason+Long mixture with Nimble added.
- Reports 73.2% mean accuracy on the dev split and 71.9% on the test split across 13 datasets.
- Requires fp32 weights; bf16 changed accuracy by up to 8 points on long documents (MuSiQue) in the authors' tests.
Questions about this entry
Short answers drawn from the linked project or page.
Use transformers with AutoTokenizer and AutoModelForMaskedLM on the repo Athekunal/ultrabert-jev-380m-sqrt-softplus, load the weights in fp32 with dtype=torch.float32, move the model to CUDA, and call eval(). It was tested with transformers 5.20.2 and torch 2.9.
The masked-token logits are large, around 200, and bf16 weights changed accuracy by up to 8 points on long documents (MuSiQue) in the authors' tests. All reported numbers use fp32 weights.
No. The model was trained on a specific prompt format using the tokens " yes" and " no"; plain fill-mask use will not give the reported results. The model card's prompt must be used exactly.
No. The base repo's remote code names the weights differently and would leave them unloaded. The UltraBERT model class is included in transformers 5.20.2 and torch 2.9.
73.2% mean accuracy on the dev split and 71.9% on the test split, averaged over the 13 datasets of the Reason+Long mixture.
Apache-2.0, fine-tuned from dmis-lab/UltraBERT-380M, with 429,966,336 F32 parameters stored in a single unsharded safetensors file.
It is an English encoder decision model, not a generator: it produces no text and expects the exact trained prompt format. For other languages or free-form output, another resource is needed.
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
Fine-tuned UltraBERT-380M decision model on Hugging Face that scores a state, question, and option list in one forward pass. Uses sigmoid of yes minus no normalized over options; 72.8% dev and 71.9% test accuracy across 13 datasets, with an evidence head documented on the card.
Jev Models & Reimplementations#open-models#arxiv#calibration
Community
arXiv paper presenting EDU-Jev, a family of single-pass System One decision models at 0.8B, 2B and 4B parameters that return a probability distribution over predefined education answer options without generating any text.
Jev Models & ReimplementationsArticle#arxiv#open-models#classification
Community
v-jev 1.4 is a Jev-style decision model on Hugging Face in 0.8B and 4B sizes, built on Qwen3.5 backbones. It returns probabilities for typed questions in one forward pass and adds value extraction and image decisions.
Jev Models & Reimplementations#open-models#classification#calibration
Community
A 7.7M-parameter decision model on Hugging Face that reads text or speech and returns one probability per option for choice, score, and noul questions. Speech runs through femto-asr, giving a 10.1M-parameter spoken pipeline.
Jev Models & Reimplementations#open-models#classification#local
Community
A Hugging Face checkpoint that fine-tunes UltraBERT-380M into a Jev-style decision model: it reads a state, a question and a list of options and picks one in a single forward pass. Reports 72.6% mean accuracy over 13 datasets on the test split.
Jev Models & Reimplementations#open-models#classification#local
Community
A 46M-parameter Russian and English decision model published on Hugging Face, tagged for classification, NLI, toxicity, moderation, and calibrated distillation. It ships a custom NQ4ForDecision class and a chat template with an optional thinking block.
Jev Models & Reimplementations#open-models#classification#calibration
Community
From 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.