CommunityJev Models & Reimplementations
Security-One 27B
Superagent's security triage model: a continual fine-tune of AutoJev-27B that returns calibrated probabilities for prompt-injection and content checks, with a release policy at 0.70 unsafe.
- Published by
- Community
- Author
- Superagent
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Self-reported BIPIA overall 99.75% and attack detection 99.83% on frozen row-identical inputs, against AutoJev's 46.00% and Jev's 71.75%.
- The card calls BIPIA a frozen binary detector adaptation over email, table, and code contexts — not the original generative leaderboard task.
- Calibration temperature 0.1453 with a release policy that flags unsafe at P(unsafe) >= 0.70.
- Trained with a rank-32 LoRA in 16 updates over 18,106 examples, with a hash audit finding zero overlaps in 1,230 protected rows.
Watch out
The repository is gated behind conditions, the validated production path is SGLang on a B200 with a 65,536-token classification context, and chat prompting does not reproduce the numbers. Its NotInject benign-only 87.61% cannot measure attack recall, and the card says prompt injection remains unsolved.
Questions about this entry
Short answers drawn from the linked project or page.
No. The card says ordinary chat-completion prompting does not reproduce the reported results; use the in-repo SGLang recipe or the hosted Security-One API. It is a decision model for triage, not a conversational assistant.
It is the card's release policy — classify unsafe when P(unsafe) is at least 0.70 — not a universal optimum. The card frames attack sensitivity versus benign tolerance as an operating point rather than a ranking, with BIPIA benign false positives at 0.50%.
As a continual fine-tune of AutoJev-27B with a rank-32 LoRA in 16 updates over 18,106 examples. A hash audit found zero overlaps in 1,230 protected rows.
SGLang on a B200 with a 65,536-token classification context. The repository is gated behind conditions, and ordinary chat prompting does not reproduce the reported numbers.
Self-reported BIPIA overall 99.75% and attack detection 99.83% on frozen row-identical inputs, against AutoJev's 46.00% and Jev's 71.75%. The card calls it a frozen binary detector adaptation over email, table and code contexts, not the original generative leaderboard task.
0.1453, with a release policy that flags unsafe when P(unsafe) is at least 0.70.
No. Its NotInject benign-only 87.61% cannot measure attack recall, and the card says prompt injection remains unsolved. Tune the threshold on your own traffic rather than trusting one number.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
YouTube: Kubernetes admission control with a decision model, fail-closed.
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
58GitHub stars
An Apache-2.0 decision engine with a TypeSafe-compatible endpoint, document answers with source passages, and a ViZDoom demo that picks actions from structured game state.
Jev Models & Reimplementations#community#python#local
Community
deepanwadhwa/OpenDecisionPython
deepanwadhwa/OpenDecisionPythonThe Hugging Face hub for Laya's open System 1 checkpoints: the 421M English model at the repo root, a 322M multilingual build in a subfolder, and the typed-decisions fine-tune, all Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
The fine-tuned Laya checkpoint for four typed-decision workflows: 0.766 accuracy on 2,000 held-out decisions, above Jev 1.13.0's published 0.727 and the teacher self-agreement ceiling it trained past.
Jev Models & Reimplementations#community#open-models#local
Community
The vision sibling: the same text-trained System 1 head on Qwen3.8-27B's image encoder, taking images zero-shot, ranked first of 20 vision decision models on the community Jev Decision Index 0.3.
Jev Models & Reimplementations#community#open-models#local
Community
Interfaze's 4B LoRA adapter over Qwen3.5-4B that speaks POST /v1/systemone: 0.807 held-out weighted accuracy and 0.061 ECE after calibration, with a 200,000-example training recipe documented.
Jev Models & Reimplementations#community#open-models#local
Community
FINAL-Bench's zero-token classifier: one forward pass plus a fitted temperature produce a full distribution over options, scoring 0.743 on the 2,000-decision typed-decisions test split.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
Interface, size, license, claims — in that order. A selection guide across the system-one catalog, from a 706K form scorer to 27B Jev students.
Jev Practices & Patterns#open-models#evaluation#benchmarks
Read article
Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.
Jev Tools & CLIs#open-models#local#onnx
Read article
Hundreds of repositories carry the system-one tag, but the ecosystem is a dozen ideas in different sizes. A tour of the families, the licenses, and what a download does not include.
Jev Models & Reimplementations#open-models#hugging-face#local
Read article
From the Jev community
Posts from builders shipping with Jev right now.