Public dataset
qev-data packages 15 historical corpus configurations, not 15 disjoint datasets.
Each archive contains original eligible JSONL records, referenced image bytes and a
publication manifest. Viewer JSONL exposes the same requests and targets as JSON strings
to avoid incompatible nested candidate schemas. Original train/dev/calibration/test values
are preserved. The release manifest records each original and exported file checksum.
Public corpus files use qev-<stage>.zip. Their extracted stage
folders and original record/source IDs preserve the recorded training provenance.
For example, qev-starter-v2.zip extracts to starter-v2/.
Attribution and licenses
| Source | License | Conversion |
|---|---|---|
| Authored diagrams, policy rules, conditional probabilities and workflows | Apache-2.0 | Generated evidence, questions and probability targets |
| Beans, AIR Lab at Makerere | MIT | Typed questions over leaf photos |
| TrashNet, Gary Thung and Mindy Yang | MIT | Recognition and request-defined policy views; Kaggle version 1 source matched by hashes |
| SNLI, Stanford NLP / Bowman et al. | CC-BY-SA-4.0 | Premise/hypothesis decisions and policy views; derived material remains share-alike |
| BANKING77, PolyAI / Casanueva et al. | CC-BY-4.0 | Sampled candidate intent and policy views |
| LocalLLaMA typed-decisions | Apache-2.0 | Recorded training partition and teacher-agreement distributions |
The dataset is a collection under multiple licenses. The repository's Apache license does not relicense CC-BY or CC-BY-SA records. Each record carries its source/license/revision; original MIT notices and CC legal texts are included. Original source observations were selected and converted into the recorded typed request format; these are modified datasets.
CIFAR-10: the upstream card marks the license unknown. Its evaluation images and requests are excluded. Exclusion manifests retain IDs, hashes, split and source references for direct upstream reconstruction. No selected training rows are removed by this license exclusion. The model repository contains no raw training corpus or third-party evaluation photos.
Integrity and use
scripts/package_public_data.py fails on duplicate IDs, group leakage, byte-identical image
leakage between splits, incorrect image hashes, paths outside a corpus or unexpected licenses.
These are checks within each corpus. They do not establish cross-stage disjointness, distinct
real objects, or absence from Qwen pretraining.
python scripts/package_public_data.py --source /path/to/original-workspace --output dist/qev-data
This maintainer command requires the recorded source snapshots. Users can download the
published archives directly; they do not need to recreate the original private workspace.
The repository includes the public dataset manifest under release/.
Original notices
See licenses/; each corpus ZIP includes these notices.
October 2026 vision mix
The vision-mix-2026-10/ folder contains no third-party images or text. For the 17 third-party sources, it holds ids, gold letters and SHA-256/dHash values only. The source table in vision-mix-2026-10/README.md gives each source and the license stated where it was downloaded. Rebuilding the items requires accepting each source's own terms. FUNSD allows non-commercial research use only, and Hateful Memes is research-only.
The synthetic set (vision-mix-2026-10/synthetic/), the NLVR2 teacher probabilities (outputs of QEV 0.2.1) and the build scripts are Apache-2.0. NLVR2 ids and labels refer to the official NLVR2 release (lil-lab/nlvr), and its terms apply to the statements and images.
October 2026 vision mix v3
vision-mix-2026-10-v3/ uses the same third-party sources as vision-mix-2026-10/, under the same terms, and adds 157 rows from the Hateful Memes and A-OKVQA train splits. It contains no third-party images or text, including the re-rendered Mind2Web screenshots and the distractor texts of regenerated items. Third-party items are described by ids, labels, hashes, generator labels and seeds only. The build scripts are Apache-2.0. The synthetic items and NLVR2 teacher probabilities it refers to are those in vision-mix-2026-10/.