| QEV (formerly Veyra) |
| Copyright 2026 QEV contributors |
| Copyright 2026 Veyra contributors |
|
|
| QEV is an independent open-source research project for dynamic typed |
| decisions over images and text. |
|
|
| The backbone is Qwen/Qwen3.5-2B, published by the Qwen team under Apache-2.0. |
| Model weights are not included in this repository and retain their upstream |
| license and attribution requirements. |
| https: |
|
|
| The design takes inspiration from LAYA's typed decision interfaces and |
| candidate scoring. No LAYA source code or model weights are included in |
| this implementation. |
| https://github.com/NandhaKishorM/laya |
| |
| TypeSafe JEV's public API is a reference for typed question and answer |
| contracts. This repository does not contain JEV source code or weights. |
| https: |
|
|
| Starter head training uses the MIT-licensed Beans dataset from AIR Lab, |
| Makerere University. Its copyright and permission notice are preserved in |
| docs/data-licenses/beans-MIT.txt. Dataset images are downloaded separately. |
| https: |
|
|
| Foundation training also uses the following separately licensed datasets. |
| Original records and photographs are not redistributed with the model package. |
|
|
| TrashNet: Gary Thung and Mindy Yang, MIT. |
| Kaggle redistribution: vminhkhoi/trashnet, version 1. Image/class byte hashes |
| were matched against the original-author dataset before split construction. |
| The original notice is preserved in docs/data-licenses/trashnet-MIT.txt. |
| https: |
| https: |
|
|
| SNLI: Stanford Natural Language Inference corpus, Stanford NLP Group, |
| Samuel R. Bowman et al. (2015), CC-BY-SA-4.0. |
| Selected premise/hypothesis pairs were converted into dynamic decision views; |
| derived dataset material retains the applicable share-alike requirements. |
| https: |
| https: |
| https: |
|
|
| BANKING77: PolyAI, Casanueva et al. (2020), CC-BY-4.0. |
| Utterances and category names were converted into sampled eight-candidate |
| questions and request-specific policy views; this is not the standard 77-way |
| evaluation. Attribution and source revisions accompany the data protocol. |
| https: |
| https: |
| https: |
|
|
| CC-BY-4.0 and CC-BY-SA-4.0 legal texts in docs/data-licenses were obtained |
| from the SPDX license-list-data text distribution. In the Hugging Face |
| package these notices are under data-licenses/ instead of docs/data-licenses/. |
|
|
| The LocalLLaMA/typed-decisions dataset is used for an external text benchmark. |
| Its Apache-2.0 teacher-agreement targets are not objective ground-truth labels. |
| https: |
|
|
| Workflow v12 additionally adapts on the recorded training partition of that |
| dataset; development, calibration and official test remain separate. Newly |
| generated workflow tasks are controlled procedural examples, not recorded |
| business outcomes. Their conditional probabilities follow their stated model. |
|
|
| CIFAR-10: Alex Krizhevsky, Vinod Nair and Geoffrey Hinton. |
| Used only as a 32-by-32 image transfer evaluation guard in Workflow v12. |
| The uoft-cs/cifar10 dataset card marks the license as unknown. No CIFAR-10 |
| images are redistributed. This does not establish permission for redistribution. |
| https: |
| https: |
|
|
| October 2026 vision mix (QEV-2B, QEV-0.8B). Only ids, labels and hashes of third-party items are included; no third-party images or text are redistributed: |
| ScienceQA (derek-thomas/ScienceQA), A-OKVQA (HuggingFaceM4/A-OKVQA) and COCO images, ChartQA (HuggingFaceM4/ChartQA), |
| CORD v2 (naver-clova-ix/cord-v2), FUNSD (Jaume et al., 2019), Hateful Memes (Kiela et al., 2020), |
| Multimodal-Mind2Web (osunlp/Multimodal-Mind2Web), and the AI2D, DocVQA, IconQA, InfographicVQA, Screen2Words, |
| TallyQA, Visual7W and VSR subsets of HuggingFaceM4/the_cauldron, plus NLVR2 (Suhr et al., 2019; lil-lab/nlvr). |
| Each remains under its own terms. |
|
|
