Jev directory
CommunityJev Models & Reimplementations
Jevvideo OneJev-4B H200 LoRA Adapters
LoRA adapter set for the Jevvideo long-video frame-selection agent, trained on 8xH200 from OmniJev/OneJev-4B. Ships controller, grounding, and verifier adapters plus per-role metrics and five-benchmark results.
- Reading time
- 2 min
- Published by
- Community
- Added
- Nov 2026
Highlights
- Three deployed LoRA adapters: controller, grounding (locator + frame selector), and verifier.
- Base model OmniJev/OneJev-4B; LoRA r=16, alpha=32, dropout 0.0 on attention and MLP layers.
- Five-benchmark mean 63.54; LVBench 54.01 versus 40.44 for uniform-32-frame sampling.
- Each adapter directory holds trainable.pt (~130 MB) and base_model.json; not plug-and-play PEFT adapters.
- Needs the Jevvideo pipeline plus a frozen Qwen3-VL-8B reader to run end to end.
Questions about this entry
Short answers drawn from the linked project or page.
The adapters are loaded by the Jevvideo codebase, which drives a window-tree navigation over a long video, and a frozen Qwen3-VL-8B reader answers from 32 frames (16 overview plus 16 kept). The reader is used off the shelf and is not included in this repository.
Base is OmniJev/OneJev-4B at revision c88e18653ceb7a8770716287f55fdefc79d6b588. LoRA is r=16, alpha=32, dropout 0.0, applied to attention (q/k/v/o, in_proj_*, out_proj) and MLP (gate/up/down_proj) across all layers.
End-to-end accuracy over 10,759 of 10,760 questions: CG-Bench mini 51.87, MLVU M-avg 76.38, LongVideoBench 66.87, Video-MME overall 68.56, LVBench 54.01, for a five-benchmark mean of 63.54. Video-MME by duration is short 79.44, medium 69.11, long 57.11.
The repository is publicly accessible but gated: you must log in or sign up and accept the conditions to access the files. The license is listed as other. It is not deployed by any Inference Provider.
adapters/{controller,grounding,verifier}/ each contain trainable.pt and base_model.json. Results live in results/<role>_metrics.jsonl, <role>_run_info.json and <role>_early_stopped.json, with configs in config/<role>_config.yaml, h200.yaml, formal_1103.yaml and version.env.
Against a uniform-32-frame baseline using the same reader, paired per question, the agent scores 54.01 versus 40.44, a +13.57-point gain with 95% CI [10.82, 16.20]. 338 questions are answered only by the agent, versus 128 only by uniform sampling.
It is a research artifact, not a general chat model, and requires the Jevvideo pipeline plus a Qwen3-VL-8B reader. Dev metrics are role-level sub-task accuracies, not end-to-end QA accuracy, so they should not be read as benchmark performance.
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-weight eligibility scorer that LoRA-adapts Qwen models with a pointer head to score natural-language answer options in one forward pass, trained on TREC Clinical Trials 2021 and evaluated on 6,975 pairs from 50 unseen synthetic patients.
Jev Models & ReimplementationsArticle#arxiv#open-models#evaluation
Community
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-jevPythonv-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 trained example bank for @ruvector/typesafe holding 9,993 labelled Banking77 utterances across 77 banking intents with frozen split assignments. Heads and temperature calibration are refit from the bank on load; no weights are stored.
Jev Models & Reimplementations#open-models#classification#local
Community
Ready-made Cascadia decision tree that serves autotrust/JEV-27B from OpenVINO int8 weights split into two pipeline stages. Grafts JEV-27B's verbalizer LoRA onto Intel's int8 Qwen3.8-27B IR so /v1/decisions and /v1/systemone run without the exporter.
Jev Models & Reimplementations#open-models#local#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.