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://huggingface.co/Qwen/Qwen3.5-2B

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://api.typesafe.ai/openapi.json

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://github.com/AI-Lab-Makerere/ibean

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://github.com/garythung/trashnet
https://www.kaggle.com/datasets/vminhkhoi/trashnet

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://nlp.stanford.edu/projects/snli/
https://huggingface.co/datasets/stanfordnlp/snli
https://creativecommons.org/licenses/by-sa/4.0/

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://github.com/PolyAI-LDN/task-specific-datasets
https://huggingface.co/datasets/PolyAI/banking77
https://creativecommons.org/licenses/by/4.0/

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://huggingface.co/datasets/LocalLLaMA/typed-decisions

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://www.cs.toronto.edu/~kriz/cifar.html
https://huggingface.co/datasets/uoft-cs/cifar10
