Jev directory
CommunityJev Models & Reimplementations
Open-Jev-9B
Zefan Cai's LoRA adapter and scalar decision head for Qwen3.5-9B score caller-supplied candidates for choice, yes/no and score answers without generating text, served over a local /v1/systemone endpoint.
- Reading time
- 2 min
- Published by
- Community
- Author
- ZefanCai
- Last verified
- Oct 2026
Used inJev for Software & DevTools
Highlights
- Needs the pinned Qwen/Qwen3.5-9B revision and the Open-Jev loader; a plain AutoPeftModel call does not implement the interface.
- Full-data evaluation: 97.54% hard accuracy on 10,532 test rows and 91.97% on 15,920 OOD rows, with ECE 0.0077 and 0.0374.
- Trained for 20,204 optimizer steps at global batch 4, one full pass over the frozen release-v2 split; LoRA rank 8, alpha 16.
- Saved temperature 1.8969 fitted on 512 calibration rows; training excluded 1,700 wikispeedia-v1 rows pending redistribution permission.
- Server rejects input over 4,096 tokens per scored candidate rather than truncating, and returns declared keys and probabilities.
Quickstart
bash
git clone https://github.com/Zefan-Cai/Open-Jev.git
cd Open-Jev
python -m pip install -e '.[train]'
hf download ZefanCai/Open-Jev-9B --local-dir ./checkpoints/open-jev-9b
python -m jev.server \
--checkpoint ./checkpoints/open-jev-9b/package/checkpoint \
--device cuda:0 --max-length 4096 --batch-size 1 --no-prefix-cache \
--host 127.0.0.1 --port 8791Watch out
Model card reviewed 2026-10-15; the fetched text was truncated inside the artifact table. The card says no full-data baseline was evaluated, so the table does not establish training gain, and the redistributable dataset is not byte-identical to the 80,816-row training set.
Questions about this entry
Short answers drawn from the linked project or page.
A LoRA adapter plus a trained scalar decision head for Qwen/Qwen3.5-9B that returns typed decisions: probabilities over caller-supplied choices, a yes/no probability, or ordinal levels with their expected value, without generating answer text.
A suitable GPU plus the upstream Qwen3.5-9B weights available locally or downloadable from Hugging Face; the loader uses the pinned upstream revision from package/checkpoint/model.json. This repository ships the decision checkpoint, not merged base weights.
Clone the Open-Jev repository, install with pip install -e '.[train]' (Transformers 5.10.2 and PEFT 0.19.1), download the checkpoint into ./checkpoints/open-jev-9b, then run python -m jev.server with --device cuda:0 on port 8791.
POST JSON to /v1/systemone with a state string and a questions map, for example an intent choice over billing, technical and other criteria. The server returns declared keys and probabilities and does not execute the proposed actions.
20,204 optimizer steps at global batch 4 over one full pass of the frozen release-v2 split. Full-data evaluation reports 97.54% hard accuracy on test and 91.97% on OOD, with a saved temperature of 1.8969 fitted on 512 calibration rows.
Maximum input is 4,096 tokens per independently scored candidate, and input over that is rejected rather than silently truncated. Prefix caching is opt-in and disabled in the documented launch command.
No. The card says no full-data baseline was evaluated, so the table does not establish training gain. The redistribution dataset also excludes 1,700 wikispeedia-v1 rows and is not byte-identical to the training set.
More Jev resources like this
Gestalt Labs' Jeff is a LoRA adapter on Qwen3-4B-Instruct-2507 that answers Jev-style choice, noul and score questions with labels and probabilities instead of prose, runs locally, and ships a train-your-own guide; on 9,730 fact-checking examples it trails Jev 1.13.0 on accuracy (0.8183 vs 0.8283).
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
HAL-X AI's 322M non-autoregressive decision model returns calibrated answers in one forward pass of about 10 ms, with native number, excerpt and citation outputs beyond Jev's choice, noul and score types. Apache-2.0, post-trained in 18 languages.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
An open-source, Jev-compatible System One model: Dex answers choice, score and yes/no questions over text in a single forward pass from an Ettin-1B encoder, and speaks the same /v1/systemone shape as Jev.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
canadaornaw/dexPython
canadaornaw/dexPythonA 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 CPU port of SemIf with a web UI: JEV-CPU loads Qwen3-0.6B in float32 and reads a decision from its option logits in about a second, across eight domains from support to incident severity, with no GPU.
Jev Models & Reimplementations#community#catalog-import#madewithjev
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.