This dataset repackages images and labels from many third-party datasets. Several of them are research-only or non-commercial (RVL-CDIP, AVA, ScienceQA, AG News, Yelp, Food-101, Oxford Flowers, Stanford Cars), and images keep their original copyright. Every record carries its source licence in the license column. By requesting access you agree to use this data for non-commercial research only and to follow the licence and terms of each source dataset.

sev-ood

Typed, labelled decisions about images, in the TypeSafe /v1/systemone request shape: an image plus a state (text or JSON), and one or more typed questions (choice, noul = yes/no, score = ordered levels), each with its label. It was built to fine-tune Kev, a Jev-style decision model, to read images (code: kev-vision), but any image classifier / VLM / reward model can use it.

sev-ood (evaluation only): unseen games, unseen classes, unknowable questions and corrupted images; never trained on.

Gated, research use only. This repository repackages third-party data. Access is granted automatically once you accept the terms: non-commercial research only, and the licence of each source (the license column) applies to its records. Images keep their original copyright. If you are a rights holder and want a source removed, open a discussion on this repository.

Record format

column	content
`image`	the image (JPEG/PNG), or null for text-only replay records
`state`	JSON: a string or an object (e.g. `{"game": ..., "previous_action": ...}`)
`questions`	JSON: `{qid: {"type", "instructions", "criteria", "label", "target"?, "src"}}`. Labels: the option name for `choice`, `true`/`false` for `noul`, the level index (from 0) for `score`. `target`, when present, is a soft distribution over the options (uniform when the image cannot answer; the voters' histogram for AVA).
`source`, `license`, `split`, `id`	provenance: every record carries the licence of the source it was derived from

No augmentation is stored: option order, "none of the above" insertion and distractors are applied at training time.

What was added, source by source

source	licence	questions	how it was converted
`ood_qbert`	Apache-2.0 (JAT)	choice	see kev_vision.ood
`ood_seaquest`	Apache-2.0 (JAT)	choice	see kev_vision.ood
`ood_boxing`	Apache-2.0 (JAT)	choice	see kev_vision.ood
`ood_beamrider`	Apache-2.0 (JAT)	choice	see kev_vision.ood
`ood_food101`	research only (source terms)	choice	see kev_vision.ood
`ood_flowers`	research only (source terms)	choice	see kev_vision.ood
`ood_cars`	research only (source terms)	choice	see kev_vision.ood
`unknowable`	derived from the test splits (see their licences)	choice	see kev_vision.ood
`corrupted_eurosat`	MIT	choice	see kev_vision.ood
`corrupted_pets`	CC-BY-SA-4.0	choice	see kev_vision.ood
`corrupted_vqav2`	CC-BY-4.0; COCO images	choice	see kev_vision.ood

Splits and counts

train / validation / test are disjoint within each source (the source's own splits where it has them; otherwise by image id or episode). test_ood holds sources that are never used for training (out-of-domain evaluation). VQAv2 and POPE both come from COCO val2014: every POPE image is excluded from VQAv2, and A-OKVQA (COCO 2017, which contains val2014) skips any photo matching a POPE or VQAv2 validation/test image by perceptual hash.

source	split	records	choice	noul	score
`ood_qbert`	test_ood	500	500	0	0
`ood_seaquest`	test_ood	500	500	0	0
`ood_boxing`	test_ood	500	500	0	0
`ood_beamrider`	test_ood	500	500	0	0
`ood_food101`	test_ood	500	500	0	0
`ood_flowers`	test_ood	500	500	0	0
`ood_cars`	test_ood	500	500	0	0
`unknowable`	test_ood	492	492	0	0
`corrupted_eurosat`	test_ood	500	500	474	0
`corrupted_pets`	test_ood	500	500	334	0
`corrupted_vqav2`	test_ood	494	0	756	217

Balance audit (train split)

majority is the accuracy of always giving the most frequent answer, i.e. what a model that ignores the image would score; copy prev is the accuracy of repeating the previous action (control records that show it). Train images shared with validation/test: none.

question	n	classes	majority	yes rate	copy prev	answer is option 1

Licences

This dataset is a derivative of the sources above; each record keeps its source's licence in the license column and you must follow it. COCO photos (VQAv2, A-OKVQA, POPE) remain under their original Flickr licences.

Downloads last month: 15

Total file size:

137 MB

Models trained or fine-tuned on Jacqkues/sev-ood

Jacqkues/sev-0.8b

Visual Question Answering • Updated 6 days ago • 20