Jev directory
CommunityJev Models & Reimplementations
open-jev DeBERTa ONNX
A Transformers.js ONNX conversion of open-jev-deberta-v3-large: fp16, fp32, q4 and q4f16 graphs from 0.35 GB that score typed questions in the browser through a single forward pass.
- Reading time
- 2 min
- Published by
- Community
- Author
- onnx-community
- Last verified
- Oct 2026
- Languages
- JavaScript
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Four variants under onnx/: fp16 at 0.88 GB as the default, fp32 1.75 GB, q4 0.48 GB and q4f16 0.35 GB.
- Runs through Transformers.js with a WebGPU device; one forward pass returns a logit per (question, option) pair.
- The graph takes token ids, attention mask, a span slot per token and the slot ids of each pair, so several typed questions share one sequence.
- Sequence limit of 512 tokens with the state cut to 256, matching the source model, and temperature 1.05 for the softmax.
- Converted from com-kotobalabs/open-jev-deberta-v3-large, whose card reports 0.854 in-domain accuracy and ECE 0.022.
Watch out
The 512-token window and 256-token state cut come straight from the source model, and the temperature 1.05 is the source's calibration, not a new fit. The ONNX card does not restate the source's limitations, but they still apply: English-only, three public domains and confidence that is over-confident out of distribution.
Questions about this entry
Short answers drawn from the linked project or page.
fp16 is the default at 0.88 GB, fp32 is 1.75 GB, q4 is 0.48 GB and q4f16 is 0.35 GB. The card gives no per-variant accuracy table, so treat the smaller files as a size trade-off to validate on your own inputs.
input_ids and attention_mask for the sequence, a seg tensor giving each token its span slot (option tokens get their pair index, question tokens offset by pair count, everything else -1), and pair_q and pair_opt slot ids for each scored pair.
One sequence carries the state and all typed questions, and the single forward pass returns a logit per question-option pair. The included JavaScript example then applies a softmax per option group to produce each question's distribution.
Import AutoModel and AutoTokenizer from @huggingface/transformers, load the repo with dtype 'fp16' and device 'webgpu', build the token and slot tensors as in the card's example, and read logits back to float32 for the softmax.
The sequence is limited to 512 tokens with the state cut to 256, exactly as in the source model. The example's softmax uses temperature 1.05, which the card identifies as the source model's calibration.
The conversion card does not publish its own numbers; the source card reports 0.854 in-domain accuracy with ECE 0.022 on 1,500 test states, dropping to 0.690 out of distribution. Validate the quantized variants before relying on those figures.
The repository is tagged apache-2.0 and converted from microsoft/deberta-v3-large, whose base is MIT, with datasets banking77 (CC-BY-4.0), SST-5 and BoolQ (CC-BY-SA-3.0). The source card's provenance section remains authoritative.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Text-to-screenshot (Malay Vasa)
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
X: Google's MediaPipe Decision Maker, tested in the browser
More 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
Mithril's open, Jev-shaped typed-decision model on DeBERTa-v3-large: choice up to 255 options, score and noul from one forward pass, 0.854 in-domain with ECE 0.022, trained on public gold labels only.
Jev Models & Reimplementations#community#open-models#local
Community
A CPU-first BitNet student: Microsoft's BitNet b1.58 2B quantized to an I2_S GGUF with a float32 pointer head, distilled from Kev 9B logits and run through the bit-jev Python package with no separate open-weight licence.
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
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.