Jev directory
CommunityJev Models & Reimplementations
Visual Jev 4B
A LoRA adapter for Qwen3-VL-4B-Instruct that scores supplied visual answer candidates instead of writing prose: macro accuracy over four choice benchmarks rises from 0.706 to 0.761, Apache-2.0.
- Reading time
- 3 min
- Published by
- Community
- Author
- guanxuyu
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- A LoRA for Qwen3-VL-4B-Instruct with the vision tower frozen: r=16, alpha=32 and 33,030,144 trainable parameters.
- Equal-weight macro over four choice benchmarks rises from 0.706 to 0.761, with a 0.002 spread over three seeds.
- GQA-Choice improves 0.880 to 0.916 and SNLI-VE 0.629 to 0.808; held-out families move only 0.001 and 0.005.
- Three seeds ship together: seed 0 at the root and seed1/ plus seed2/ subfolders, all under Apache-2.0.
- Trained for 3,000 steps on GQA and SNLI-VE records only, with K in {2, 4} candidates and no new human annotation.
Watch out
Apache-2.0 adapter over a separately licensed Qwen3-VL-4B-Instruct backbone. All gains sit on the two training task families: held-out TextVQA and TallyQA move 0.001 and 0.005 against a 0.014 seed spread, and TallyQA's 0.345 is weak in absolute terms. Scores require the fixed readout position and exact prompt layout, and the reported probabilities are bf16 arithmetic, not calibrated evidence.
Questions about this entry
Short answers drawn from the linked project or page.
A score per supplied candidate answer. One image and one shared text context drive a set of questions whose options you pass at request time; the model reads candidate-token logits from the backbone's own LM head at one fixed readout position and does not write prose.
Macro accuracy rises from 0.706 to 0.761 plus or minus 0.002 over three seeds. GQA-Choice goes 0.880 to 0.916 and SNLI-VE 0.629 to 0.808, while the held-out TextVQA-Choice and TallyQA-Choice families move only 0.974 to 0.975 and 0.340 to 0.345.
Seed 0 is at the repository root (adapter_config.json and adapter_model.safetensors); the other two are the subfolders seed1/ and seed2/. Load them by passing subfolder="seed1" to PeftModel.from_pretrained over the Qwen/Qwen3-VL-4B-Instruct base.
Not reliably. The numbers come from candidate-token logits at one fixed readout position with a specific prompt layout and candidate tokenisation, so a custom prompt measures a different setup; the linked code repository carries the exact construction, inference path and reproduction guide.
Training records derive from GQA and SNLI-VE training splits only, with GQA Choice items presented as two or four options. TextVQA-Choice and TallyQA-Choice were never seen in training, which the card marks with a dagger in its results table.
Choice mode normalizes over the options you supply and has no built-in abstention, so add an option such as incorrect_question or not enough evidence. The quickstart's incorrect_question choice caught a dog/person mismatch at 0.966, a score relative to the supplied choices rather than a calibrated evidence measure.
It is an Apache-2.0 LoRA delta over Qwen/Qwen3-VL-4B-Instruct, whose backbone must be obtained under its own licence. No images are redistributed, and the card traces path-to-path bf16 logit deviations to arithmetic that collapses by four orders of magnitude in float32.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: A one-token logprob wrapper that also classifies webcam frames.
allanrbo151 pointsA single function Jev-like wrapper for LLMs, including vision modelsRead the thread on Hacker NewsX: Happy JevDay: images, box, longer context, and an opt-in to find jaggedness
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
More Jev resources like this
chaoliangUNSW's v1 Jev-Style decision model in GGUF from 1.3 GB: calibrated choice, bool and score answers on Qwen3.5-2B-Base, 82.3% on five held-out task families with ECE 0.017.
Jev Models & Reimplementations#community#open-models#local
Community
An Apple Silicon MLX bf16 build of the v1 Jev-Style decision model: 77 ms decisions on an M1 Max with calibrated choice, bool and score probabilities, Apache-2.0 weights.
Jev Models & Reimplementations#community#open-models#local
Community
Jev-compatible decision models with vision input: LoRA fine-tunes of Qwen3.5 at 4B and 9B serving Choice, Score and Noul on the /v1/systemone wire format, with screenshots and photos placed beside text in the state. Runs on your own NVIDIA GPU; weights are non-commercial.
Jev Models & Reimplementations#community#vision#open-weights
Community
Xiaooolong/vevPython
Xiaooolong/vevPythonA LoRA plus 24-slot decision head over Qwen3.8-27B from Ming Xu's train-jev project: it serves TypeSafe's System One API with 0.851 Kev transfer-v4 accuracy and ECE 0.053, and needs an 80 GB GPU.
Jev Models & Reimplementations#community#open-models#local
Community
A 24 GB-GPU alternative from the train-jev project: Qwen3.5-4B plus a rank-16 LoRA and 24-slot decision head scoring 0.647 PK accuracy, 40.4 Decision Index and 43 ms p50 latency, ahead of Kev-4B at the same size.
Jev Models & Reimplementations#community#open-models#local
Community
A Laya fine-tune for coding decisions: roughly 9k rows from SWE-Gym teach keep/drop file relevance, priority scoring and injection detection, reporting 0.833 accuracy on 1,150 held-out rows at 126 ms p50.
Jev Models & Reimplementations#community#open-models#local
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.