Jev directory
CommunityJev Models & Reimplementations
Medical Jev 4B Full
Full-parameter fine-tune of Qwen3.5-4B for medical multiple-choice scoring. Given a case and candidate options, it returns the chosen option key with per-option probabilities as short JSON, without explanations or chain-of-thought.
- Reading time
- 2 min
- Published by
- Community
- Added
- Nov 2026
Highlights
- Full-parameter fine-tune of Qwen3.5-4B; all roughly 4.2 billion parameters updated on 44,847 training samples.
- Each call performs one forward scoring pass and returns a short JSON answer with probabilities, no explanations.
- Development set of 2,001 questions scores 83.21%; external benchmark of 1,514 questions scores 63.21%.
- Supports 2 to 52 candidates and full inputs up to 8,192 tokens including material, options, and template.
- Download is about 8.4 GB of weights plus a roughly 9.3 GB pinned Qwen3.5-4B base; FP16 by default.
Questions about this entry
Short answers drawn from the linked project or page.
Create a virtual environment, install torch 2.8.0 from the cu128 index and huggingface-hub 1.33.0, download the bundle at revision v1.0.0, install its requirements.txt, then run predict_choices.py with --bundle pointing at the download and --input pointing at a JSONL file.
Weights use a custom jev-full-v1 format and must be loaded with the bundled predict_choices.py or ChoicePredictor. It is not a PEFT adapter and not a chat generation interface.
The download package contains about 8.4 GB of full weights plus standalone inference code and examples. First load also fetches a pinned Qwen3.5-4B base of about 9.3 GB, which can be reused with --base-model if you already have it.
The LoRA version scores 83.81% on the 2,001-question development set and 67.90% on the 1,514-question external benchmark; the full fine-tune scores 83.21% and 63.21%. The external gap of 4.69 points has a grouped bootstrap 95% interval of -6.55 to -2.90 points, and the LoRA version is currently recommended.
Input is JSONL with one question per line, containing a state with the material, questions with instructions, and criteria mapping option keys to labels. Output includes the id, a status, the answer key, and a probabilities object per option, for example 0.972 for the selected option.
The model card lists an apache-2.0 license and tags for Chinese and English, along with medical, choice-scoring, full-finetuning, and qwen3_5.
It supports 2 to 52 candidates and full inputs up to 8,192 tokens. A set of 1,620 sealed test questions has not yet been evaluated, and the documented scenarios are offline prototyping, teaching drills, and model evaluation rather than clinical use.
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
A preregistered measurement of a Jev-style interface on frozen Qwen3-4B: read the option letter's logits instead of generating grammar-constrained JSON.
Jev Models & Reimplementations#community#python#local
Community
r-ms/mini-jevPython
r-ms/mini-jevPythonRussian-first open typed-decision model on Hugging Face. Given a context and a set of candidate options it returns a calibrated probability and picks one, covering yes/no, choice, and ordinal scale decisions. Built on mmBERT-base with 307M parameters under an MIT license.
Jev Models & Reimplementations#open-models#calibration#classification
Community
Hugging Face model repo hosting llm-as-jev, a fine-tune of Qwen/Qwen3-0.6B tagged for decision-model and multiple-choice work. Ships safetensors weights, a transformers configuration, and an Apache-2.0 license.
Jev Models & Reimplementations#open-models#local#classification
Community
A JEV decision model built on inclusionAI/Ling-3.0-tiny with AReno, shipping a decision head and rank-128 backbone LoRA adapters. Serves choice, noul, and score questions over an areno /v1/systemone endpoint on Apple Silicon MLX and CUDA.
Jev Models & Reimplementations#open-models#local#python
Community
Hugging Face text-classification checkpoint from Sagicc: a 4.9M-parameter F32 safetensors model with custom code, MIT licensed, documented with JEV decision APIs through AutoModel and a FastAPI server.
Jev Models & Reimplementations#open-models#python#local
Community
A native MLX runtime for Laya's open typed-decision weights: short decisions on Apple Silicon, with no generated tokens, PyTorch, or cloud API.
Jev Models & Reimplementations#community#python#open-models
Community
mizorewww/laya-mlxPython
mizorewww/laya-mlxPythonFrom 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.