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.

kev-vision-decisions-full

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.

Full tier: includes research-only / non-commercial sources (see the licence table).

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
`vqav2`	CC-BY-4.0 (annotations); COCO images under their Flickr licenses	noul, score	yes/no and counting questions of VQAv2 (val2014 annotations) with ≥ 8/10 annotators agreeing; counts in 6 levels (zero … five or more); up to 4 questions per image; own split by COCO id % 10
`aokvqa`	Apache-2.0 (annotations); COCO images	choice	A-OKVQA multiple choice, 4 options
`eurosat`	MIT	choice, noul	land use among 4–6 sampled classes with descriptions; "is this ?" balanced 50/50
`pets`	CC-BY-SA-4.0	choice, noul	breed among 5 sampled breeds; cat or dog
`fairface_age`	CC-BY-4.0	score	apparent age in 9 ordered bins (gender and ethnicity deliberately not asked)
`chartnet`	CDLA-Permissive-2.0	choice, noul	chart type among 5 sampled types (ChartNet core_permissive); "is this a ?"
`documents`	CDLA-Permissive-1.0 (DocLayNet) / CC-BY-4.0 (CORD) / MIT (invoices)	choice, noul	document page type among 5 of 8 classes: DocLayNet's financial reports, scientific articles, laws, government tenders, manuals, patents (balanced), CORD receipts, invoices; "is this a ?"
`pong`	Apache-2.0 (JAT)	choice	expert (JAT) next action; 4 consecutive frames tiled left→right, previous action in the state; ALE actions with the same effect merged
`breakout`	Apache-2.0 (JAT)	choice	same as Pong: stay / launch / right / left
`freeway`	Apache-2.0 (JAT)	choice	same as Pong: wait / up / down
`spaceinvaders`	Apache-2.0 (JAT)	choice	same as Pong: 6 actions (move × fire)
`mspacman`	Apache-2.0 (JAT)	choice	same as Pong: 9 joystick directions
`pusht`	MIT	choice, score	LeRobot PushT human demos: direction of the next end-effector target (8 compass sectors or hold) and step size (3 levels); previous move in the state
`text`	CC-BY-4.0 / CC-BY-SA-3.0 / MultiNLI terms	choice, noul, score	text-only replay from Kev's training sources (banking77, BoolQ, MultiNLI), no image
`blank`	CC0-1.0 (generated)	choice	an empty image of a random colour with a question it cannot answer; uniform soft target (teaches "I can't tell")
`pope`	COCO images; POPE annotations (MIT)	noul	test_ood only: POPE object-presence questions (random / popular / adversarial), balanced
`scienceqa`	CC-BY-NC-SA-4.0	choice	ScienceQA questions that have an image; the hint passage goes in the state
`rvl_cdip`	IIT-CDIP terms (research only)	choice	document type among 6 of the 16 RVL-CDIP classes
`ava`	DPChallenge photos (research only)	score	aesthetic score 1–10; soft target = the voters' histogram (≥ 50 votes)
`text_full`	AG News (non-commercial), SST, Yelp dataset terms	choice, noul, score	text-only replay from AG News, SST-5, Yelp
`countbench`	unspecified (LAION images)	score	test_ood only: CountBenchQA counting, 11 levels
`ai2d`	unspecified (AI2 diagrams)	choice	test_ood only: AI2D science-diagram multiple choice

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
`vqav2`	test	199	0	318	83
`vqav2`	train	24782	0	40749	9802
`vqav2`	validation	1249	0	2010	516
`aokvqa`	test	200	200	0	0
`aokvqa`	train	11999	11999	0	0
`aokvqa`	validation	518	518	0	0
`eurosat`	test	200	200	198	0
`eurosat`	train	9000	9000	8992	0
`eurosat`	validation	450	450	432	0
`pets`	test	200	200	124	0
`pets`	train	3500	3500	2274	0
`pets`	validation	175	175	130	0
`fairface_age`	test	200	0	0	200
`fairface_age`	train	7500	0	0	7500
`fairface_age`	validation	375	0	0	375
`chartnet`	test	200	200	182	0
`chartnet`	train	9000	9000	8990	0
`chartnet`	validation	450	450	444	0
`documents`	test	200	200	188	0
`documents`	train	10225	10225	10056	0
`documents`	validation	575	575	570	0
`pong`	test	158	158	0	0
`pong`	train	1500	1500	0	0
`pong`	validation	64	64	0	0
`breakout`	test	200	200	0	0
`breakout`	train	1500	1500	0	0
`breakout`	validation	46	46	0	0
`freeway`	test	188	188	0	0
`freeway`	train	1500	1500	0	0
`freeway`	validation	72	72	0	0
`spaceinvaders`	test	200	200	0	0
`spaceinvaders`	train	1500	1500	0	0
`spaceinvaders`	validation	55	55	0	0
`mspacman`	test	179	179	0	0
`mspacman`	train	1500	1500	0	0
`mspacman`	validation	69	69	0	0
`pusht`	test	200	200	0	185
`pusht`	train	6000	6000	0	5462
`pusht`	validation	300	300	0	277
`text`	test	200	133	67	0
`text`	train	7998	5332	2666	0
`text`	validation	400	262	138	0
`blank`	test	200	200	0	0
`blank`	train	4500	4500	0	0
`blank`	validation	225	225	0	0
`pope`	test_ood	486	0	486	0
`scienceqa`	test	139	139	0	0
`scienceqa`	train	6000	6000	0	0
`scienceqa`	validation	197	197	0	0
`rvl_cdip`	test	180	180	0	0
`rvl_cdip`	train	4000	4000	0	0
`rvl_cdip`	validation	180	180	0	0
`ava`	test	200	0	0	200
`ava`	train	6000	0	0	6000
`ava`	validation	300	0	0	300
`text_full`	test	200	64	197	136
`text_full`	train	4998	1666	4998	3332
`text_full`	validation	250	79	247	171
`countbench`	test_ood	491	0	0	491
`ai2d`	test_ood	500	500	0	0

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
`agnews`	1666	4	0.256			0.238
`agnews_yn`	3332	2	0.755	0.245
`aokvqa`	11999	4	0.253			0.254
`ava`	6000	9	0.563
`banking77`	2666	77	0.022			0.013
`blank`	4500	10	0.105			0.232
`boolq`	2666	2	0.613	0.613
`breakout_action`	1500	4	0.359		0.476	0.275
`chartnet_type`	9000	20	0.092			0.207
`chartnet_yn`	8990	2	0.5	0.5
`documents_type`	10225	8	0.147			0.197
`documents_yn`	10056	2	0.5	0.5
`eurosat`	9000	10	0.116			0.203
`eurosat_yn`	8992	2	0.5	0.5
`fairface_age`	7500	9	0.3
`freeway_action`	1500	3	0.629		0.689	0.342
`mnli`	2666	3	0.344			0.336
`mspacman_action`	1500	9	0.15		0.571	0.095
`pets_breed`	3500	37	0.028			0.199
`pets_dog`	2274	2	0.5	0.5
`pong_action`	1500	3	0.443		0.473	0.334
`pusht_direction`	6000	9	0.127		0.745	0.11
`pusht_step`	5462	3	0.415
`rvl_cdip`	4000	16	0.089			0.165
`scienceqa`	6000	5	0.362			0.367
`spaceinvaders_action`	1500	6	0.199		0.583	0.171
`sst5`	1666	5	0.269
`vqav2_count`	9802	6	0.289
`vqav2_yn`	40749	2	0.5	0.5
`yelp`	1666	5	0.207
`yelp_yn`	1666	2	0.613	0.387

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: 25

Total file size:

4.25 GB

Models trained or fine-tuned on Jacqkues/kev-vision-decisions-full

Jacqkues/sev-0.8b

Visual Question Answering • Updated 2 days ago • 17