Jev directory
CommunityJev Models & Reimplementations
Train-Jev 4B LoRA
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.
- Reading time
- 2 min
- Published by
- Community
- Author
- shibing624
- Last verified
- Oct 2026
Used inJev for Software & DevTools
Highlights
- Qwen3.5-4B with a rank-16 LoRA on attention, DeltaNet and MLP projections plus a 24-slot fp32 head.
- Ahead of Kev-4B at the same size: 13-dataset PK accuracy 0.647 vs 0.640, Decision Index 40.4 vs 38.4.
- Lowest PK calibration error of the compared models at ECE 0.076; Kev transfer-v4 accuracy 0.779.
- Runs on any 24 GB GPU with 43 ms p50 per request on an H20.
- Same 5,123-step batch-128 recipe as the 27B variant, trained on SargeDev/jev-distill-corpus-v3.
Quickstart
bash
git clone https://github.com/shibing624/train-jev.git && cd train-jev
pip install -r requirements.txt
python -m trainjev.serve --base Qwen/Qwen3.5-4B --adapter shibing624/train-jev-4b-lora --port 8020Watch out
Card reviewed 2026-10-17; the benchmark table is self-reported by the author's harness, calibration is in-distribution only, only the first 16 choice labels A-P are trained, the zero-shot image figure (0.697) sits well below the 27B variant, and everything comes from a single seed without third-party reproduction.
Questions about this entry
Short answers drawn from the linked project or page.
A JEV-style System 1 decision model: Qwen3.5-4B plus a LoRA adapter, a 24-slot decision head and per-type temperatures. It answers yes/no, multiple-choice and rating questions with calibrated probabilities in one forward pass and never generates text.
The card reports it ahead of Kev-4B at the same size: 13-dataset PK accuracy 0.647 vs 0.640 and Decision Index 40.4 vs 38.4, with the lowest PK calibration error among the compared models (ECE 0.076). Kev-4B keeps the edge on Kev transfer-v4 (0.817 vs 0.779).
Any 24 GB GPU, per the card, at 43 ms p50 per request on an H20. The bigger train-jev-27b-lora variant needs 80 GB, so this is the small, fast option in the pair.
Clone github.com/shibing624/train-jev, install requirements and run python -m trainjev.serve with base Qwen/Qwen3.5-4B and adapter shibing624/train-jev-4b-lora on port 8020; the server exposes TypeSafe's System One API at /v1/systemone.
The LoRA adapter, head.pt for the 24-slot decision head and calibration.json for the fitted temperatures. The backbone is downloaded on first start, and you must load it through train-jev because plain PEFT or Transformers loading skips the head and temperatures.
13-dataset PK accuracy 0.647 with ECE 0.076, Kev transfer-v4 accuracy 0.779 with ECE 0.075, Decision Index 40.4, zero-shot image decisions 0.697 and 43 ms p50 latency. All figures are self-reported in the card's comparison table.
Ming Xu's independent research reproduction (train-jev), Apache-2.0, and explicitly not affiliated with TypeSafe AI or autotrust. Labels come from the SargeDev/jev-distill-corpus-v3 distillation corpus, and limitations list single-seed training and only the A-P choice labels.
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 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
Mapika's NVFP4 fine-tune of Gemma-4-31B-it answers typed Choice, Noul and Score questions with calibrated probabilities and no generated text, self-reporting a 64.27 chance-corrected public Decision Index and 39 ms median latency on a B300.
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
A Laya fine-tune from Convai Innovations on the 1,200-case typed-decisions benchmark: it reports 0.777 accuracy on the official 400-case test set, above TypeSafe Jev 1.13.0 at 0.727 and the teacher ceiling at 0.735.
Jev Models & Reimplementations#community#open-models#local
Community
A 9B bilingual Turkish and English Jev-class decision model from Berkay Hasret: a 173 MB LoRA adapter on Qwen/Qwen3.5-9B that answers up to 16 typed questions about one text in a single forward pass and reports 0.783 macro accuracy on full JevBench.
Jev Models & Reimplementations#community#open-models#local
Community
A bilingual Turkish and English 4B Jev-class decision model from Berkay Hasret: a roughly 130 MB LoRA adapter on Qwen/Qwen3.5-4B that answers up to 16 typed questions per text in one pass and reports 0.774 JevBench macro accuracy.
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.