Jev-Style 2B Decision v4 — ONNX / WebGPU

The repository is artifact-only: it holds graphs, weights, tokenizer and processor metadata, example inputs, provenance and validation. A dedicated browser runner must implement Jev rendering, block attention, image preparation and verdict readout. The browser application and converter source are not included here.

No. The model is published in a multi-session ONNX format that pipeline() and AutoModel.from_pretrained() cannot directly load. It is intended for a browser runtime that handles the separate sessions and the Jev rendering and readout steps itself.

Editable text or JSON state, choice questions with named options and descriptions, true/false questions with optional criteria, score questions with 2 to 10 ordered level descriptions, and vision questions with one selected image referenced as <|image_1|> exactly once. Multiple questions are evaluated sequentially over the same state.

Up to 2,048 total tokens including state, image tokens, question and options. Choose the smallest context bucket that fits: 512, 1,024 or 2,048. Requests exceeding 2,048 tokens must be rejected without truncation. One image is allowed, on a fixed 416 x 608 canvas with aspect ratio preserved.

No. It scores supplied options and returns a verdict; it does not generate text or bounding boxes. Box questions are not implemented, and neither are multiple images, dynamic shapes or prefix-cache reuse.

It is Apache-2.0 licensed and converted from chaoliangUNSW/Jev-Style-2B-Decision-v4 at immutable revision 61e43347cd7b5342e1b86d85787e9c71112166e0. It is an independent, experimental conversion rather than an upstream release.

The original validated five prepared 64-token text examples and one 320-token receipt example. This conversion adds decision_512, decision_1024 and decision_2048 entry points for new inputs within budget, while legacy graphs remain for reproducibility. It is not the original's 32,768-token context or flexible smart-resize pipeline.