Jev directory
CommunityJev Models & Reimplementations
J33-Jev
J33.AI's small System One decision model: an EmbeddingGemma 2 encoder plus a 0.59M scorer, 271.6M parameters total, reporting 75.2% accuracy and ECE 0.019 on 2,000 typed-decisions test decisions while deciding alone on 39.9%.
- Reading time
- 2 min
- Published by
- Community
- Author
- J33 AI
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- EmbeddingGemma 2 text encoder with a 0.59M scorer, 271.6M parameters overall and no extra transformer head on top.
- Self-reported 75.2% accuracy and ECE 0.019 on the typed-decisions test split of 400 cases and 2,000 decisions.
- At its plus 1 / minus 5 scoring the model decides alone only above 83.3% confidence, doing so on 39.9% of decisions.
- One temperature per question type, fitted by LBFGS: choice 0.467, score 0.455, noul 0.663.
- Median 33 ms per case in fp32 on the training GPU; state is truncated to 4,000 characters.
Quickstart
python
from huggingface_hub import snapshot_download
import sys
path = snapshot_download("J33-AI/j33-jev")
sys.path.insert(0, path)
import gemma_jev
model = gemma_jev.load(f"{path}/model.safetensors", base="google/embeddinggemma-2",
vision=False, head="none", device="cpu")Watch out
Card reviewed 2026-10-17; the accuracy, ECE and work-saved figures are self-reported, and the card warns this model and the fitted Laya were trained on the benchmark's own train split while hosted Jev ran as delivered. Gold labels are an AI labeller's averaged samples with 73.5% self-agreement, each variant is a single seed-0 run, and it is English-only and text-only.
Questions about this entry
Short answers drawn from the linked project or page.
A small System One decision model from J33.AI: you give it a case such as a ticket, invoice, alert or agent trace plus a few fixed questions, and it picks an answer for each with a calibrated probability.
It is explicitly Jev-style: typed questions in, calibrated probabilities out, with choice, score and noul question types. One sequence runs per question, a mask marker sits in front of every option, and the score is read at each marker.
Google's EmbeddingGemma 2 text encoder with the vision tower off plus a 0.59M scorer, 271.6M parameters total and no extra transformer layer; it is fine-tuned on the train split of LocalLLaMA/typed-decisions.
On 400 test cases and 2,000 decisions it reports 75.2% accuracy at ECE 0.019, deciding alone on 39.9% of decisions for plus 22.5 net work saved; at its plus 1 / minus 5 scoring the break-even confidence is 83.3%.
Download the repository with snapshot_download, load model.safetensors through the bundled gemma_jev module on CPU or GPU, and call predict with a JSON-serialisable state and questions; state is truncated to 4,000 characters.
It notes this model and the fitted Laya were trained on the benchmark's own train split while hosted Jev ran as delivered, that gold labels are an AI labeller's averaged samples with 73.5% self-agreement, and that each variant is a single run at seed 0.
It is English-only and text-only, trained and tested on synthetic cases from four workflows (customer service, invoice processing, security incidents and agent-trace observability), and Apache-2.0 licensed like the base model and training data.
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 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
AutoTrust's combined System 1 and System 2 bundle on Gemma-4-26B-A4B-it: Decision Index 0.2.1 of 58.05, a 24-slot head with noul, choice and score, and the base chat model left bit-identical.
Jev Models & Reimplementations#community#open-models#local
Community
A 70.8M-parameter DeBERTa-v3-xsmall cross-encoder that answers typed decision questions in one batched pass with zero generated tokens, shipped with a calibration.json that the engine requires.
Jev Models & Reimplementations#community#open-models#local
Community
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
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.