Jev directory
CommunityJev Models & Reimplementations
PolicyLM-1.7B
Musubi's Apache-2.0 1.7B moderation classifier: it scores a message against your own policy rules in explicit mode or NVIDIA Aegis 2.0's 23 categories in aegis mode, in one pass and 19 languages.
- Reading time
- 3 min
- Published by
- Community
- Author
- Musubi
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- 1.7B open-weights moderation classifier returning a 0-1 score per category, up to 16 categories per policy; policy and message share 2,048 tokens.
- explicit mode scores short rules written at inference time with no retraining; aegis mode uses NVIDIA Aegis 2.0's 23 categories plus a Prompt harmful score.
- Median 35 ms per short chat message on one 24 GB NVIDIA L4 with six categories (22 ms on an H100); it also runs on a laptop CPU or Apple silicon.
- 0.842 accuracy on the card's custom-policy benchmark at 22 ms per message, against 0.909 for gpt-oss-safeguard-20B at 349 ms.
- Evaluated on messages in 19 languages; self-contained with no trust_remote_code, no second download, and works offline under Apache-2.0.
Watch out
Apache-2.0 weights fine-tuned from BidirLM-1.7B-Embedding and pinned to transformers 4.57.6 at revision v1.2; the card sets inference: false, so it is a download-and-run model. The benchmark table is the card's own: the gpt-oss-safeguard timing is one request listing violated categories while PolicyLM used six, and every evaluation set except OR-Bench was also used during development.
Questions about this entry
Short answers drawn from the linked project or page.
It is an open-weights moderation classifier from Musubi for places that need a decision right away, such as live chat, comments and user posts. You give it a message and a policy and it returns a score from 0 to 1 for each category of that policy.
explicit mode takes your own categories as short violation rules, not-violation rules and exception overrides written at inference time with no retraining; aegis mode uses a built-in 23-category taxonomy from NVIDIA's Aegis 2.0 for generic safety screening.
Install transformers==4.57.6, torch>=2.5.1, safetensors and huggingface_hub, then download revision v1.2 with hf download and run inference/quickstart.py; the helper is imported from the repo's inference folder and classify_batch scores a list of messages.
A message is flagged when a category score reaches the cutoff. Presets default to precision (0.335 explicit, 0.69 aegis) for rare violations and balanced (0.275, 0.45) when misses cost more; a per-call or per-category cutoff can override the preset.
The card lists child-safety enforcement, being a sole self-harm safeguard, a security boundary against adversarial users, and moderating assistant responses as out of scope, and says it cannot give a written reason for appeals, bans or takedowns.
1.7B parameters, a median 35 ms per short chat message on one 24 GB NVIDIA L4 with up to six categories and 22 ms on an H100. It also runs on a laptop CPU or Apple silicon, and can be called message by message or in batches.
On the card's custom-policy benchmark it scored 0.842 accuracy with 0.528 of policy edits followed, against gpt-oss-safeguard-20B's 0.909 and 0.716, but at 22 ms per message versus 349 ms; the card says PolicyLM beats every model under 20B it ran on that benchmark.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Both Jev 1.13 and 1.14 score 29 of 30 on 30 Turkish buyer and seller messages
YouTube: Building a contextual brand-safety scorer on Jev from scratch.
X: A website slop detector that checks 35 tells in 243 ms
X: Breaking, golden nugget, or slop
More Jev resources like this
The 322M mmBERT-base build of Laya for 100+ languages: state plus typed questions return typed answers with probabilities in one forward pass, with up to 8,192 tokens for long documents.
Jev Models & Reimplementations#community#open-models#local
Community
Three open-weights 4B decision models for English (Deem-4B), Turkish (Karar-4B) and German (Wähler-4B) that choose between options and return a probability for each. The server implements the Jev API, POST /v1/systemone, so existing Jev clients work by changing the endpoint.
Jev Models & Reimplementations#community#github#python
Community
mertkayacs/jevaltPython
mertkayacs/jevaltPythonA 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
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%.
Jev Models & Reimplementations#community#open-models#calibration
Community
InternLM's multimodal structured decision model, fine-tuned from Qwen/Qwen3.5-4B: a shared state, a schema of named questions and optional images go in, and every question returns an answer distribution from one forward pass. The card reports a 90.02 seven-suite average.
Jev Models & Reimplementations#community#open-models#multimodal
Community
InternLM's 0.8B multimodal structured decision model, fine-tuned from Qwen/Qwen3.5-0.8B: a state, named questions and optional images produce an answer distribution per question in one forward pass. Self-reported seven-suite average 79.38, Brier 0.530.
Jev Models & Reimplementations#community#open-models#multimodal
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.