Datasets:

obadakr
/

a11ydebt-stats

dataset stringclasses 1 value	screen_id stringlengths 1 5	episode_id int64	step_index int32	shard int16	in_paper_sample bool 1 class	n_nodes int32 2 2.3k	n_included int32 0 528	needs_label_candidates int32 0 239	needs_label_labelled int32 0 63	needs_label_unlabelled int32 0 239	needs_label_coverage float64 0 1 ⌀	icon_like_candidates int32 0 57	icon_like_labelled int32 0 37	icon_like_unlabelled int32 0 56	icon_like_coverage float64 0 1 ⌀
rico	0	null	null	null	true	21	7	0	0	0	null	0	0	0	null
rico	1	null	null	null	true	21	16	1	0	1	0	0	0	0	null
rico	2	null	null	null	true	35	19	3	2	1	0.666667	2	2	0	1
rico	3	null	null	null	true	75	19	3	3	0	1	3	3	0	1
rico	4	null	null	null	true	70	53	8	2	6	0.25	1	1	0	1
rico	5	null	null	null	true	58	23	4	4	0	1	3	3	0	1
rico	6	null	null	null	true	87	72	15	3	12	0.2	0	0	0	null
rico	7	null	null	null	true	94	71	11	1	10	0.090909	1	1	0	1
rico	8	null	null	null	true	46	43	2	1	1	0.5	1	1	0	1
rico	9	null	null	null	true	22	18	3	0	3	0	0	0	0	null
rico	10	null	null	null	true	111	103	26	1	25	0.038462	0	0	0	null
rico	11	null	null	null	true	100	77	17	1	16	0.058824	7	1	6	0.142857
rico	12	null	null	null	true	9	8	3	1	2	0.333333	1	1	0	1
rico	13	null	null	null	true	176	64	8	1	7	0.125	5	1	4	0.2
rico	14	null	null	null	true	112	83	19	1	18	0.052632	7	1	6	0.142857
rico	15	null	null	null	true	116	87	20	1	19	0.05	8	1	7	0.125
rico	16	null	null	null	true	22	19	4	1	3	0.25	0	0	0	null
rico	17	null	null	null	true	22	15	0	0	0	null	0	0	0	null
rico	18	null	null	null	true	29	24	4	0	4	0	1	0	1	0
rico	19	null	null	null	true	106	19	4	2	2	0.5	2	2	0	1
rico	20	null	null	null	true	9	7	2	0	2	0	0	0	0	null
rico	21	null	null	null	true	28	23	3	0	3	0	0	0	0	null
rico	22	null	null	null	true	29	25	4	0	4	0	1	0	1	0
rico	23	null	null	null	true	21	14	0	0	0	null	0	0	0	null
rico	24	null	null	null	true	7	6	1	0	1	0	1	0	1	0
rico	25	null	null	null	true	106	20	4	2	2	0.5	2	2	0	1
rico	26	null	null	null	true	50	32	7	2	5	0.285714	1	0	1	0
rico	27	null	null	null	true	81	68	2	0	2	0	0	0	0	null
rico	28	null	null	null	true	82	43	7	0	7	0	0	0	0	null
rico	29	null	null	null	true	64	48	11	0	11	0	0	0	0	null
rico	30	null	null	null	true	25	19	0	0	0	null	0	0	0	null
rico	31	null	null	null	true	136	18	3	2	1	0.666667	2	2	0	1
rico	32	null	null	null	true	163	54	19	0	19	0	5	0	5	0
rico	33	null	null	null	true	159	49	3	0	3	0	1	0	1	0
rico	34	null	null	null	true	163	107	29	0	29	0	6	0	6	0
rico	35	null	null	null	true	85	73	20	0	20	0	6	0	6	0
rico	36	null	null	null	true	6	5	0	0	0	null	0	0	0	null
rico	37	null	null	null	true	19	16	5	0	5	0	4	0	4	0
rico	38	null	null	null	true	45	25	10	0	10	0	9	0	9	0
rico	39	null	null	null	true	9	8	3	1	2	0.333333	1	1	0	1
rico	40	null	null	null	true	30	22	1	1	0	1	0	0	0	null
rico	41	null	null	null	true	30	24	2	1	1	0.5	0	0	0	null
rico	42	null	null	null	true	30	24	1	1	0	1	0	0	0	null
rico	43	null	null	null	true	20	17	1	0	1	0	0	0	0	null
rico	44	null	null	null	true	7	6	1	0	1	0	0	0	0	null
rico	45	null	null	null	true	19	16	1	0	1	0	0	0	0	null
rico	46	null	null	null	true	6	5	0	0	0	null	0	0	0	null
rico	47	null	null	null	true	22	19	2	0	2	0	0	0	0	null
rico	48	null	null	null	true	20	17	1	0	1	0	0	0	0	null
rico	49	null	null	null	true	14	11	1	0	1	0	0	0	0	null
rico	50	null	null	null	true	37	29	4	3	1	0.75	0	0	0	null
rico	51	null	null	null	true	24	18	0	0	0	null	0	0	0	null
rico	52	null	null	null	true	33	16	2	0	2	0	0	0	0	null
rico	53	null	null	null	true	94	59	5	0	5	0	0	0	0	null
rico	54	null	null	null	true	46	25	4	0	4	0	0	0	0	null
rico	55	null	null	null	true	64	19	4	2	2	0.5	2	2	0	1
rico	56	null	null	null	true	64	18	3	2	1	0.666667	2	2	0	1
rico	57	null	null	null	true	28	20	3	2	1	0.666667	0	0	0	null
rico	58	null	null	null	true	156	48	4	0	4	0	2	0	2	0
rico	59	null	null	null	true	127	33	5	0	5	0	2	0	2	0
rico	60	null	null	null	true	153	48	10	1	9	0.1	2	0	2	0
rico	61	null	null	null	true	147	52	5	0	5	0	2	0	2	0
rico	62	null	null	null	true	30	27	3	0	3	0	0	0	0	null
rico	63	null	null	null	true	36	25	8	0	8	0	6	0	6	0
rico	64	null	null	null	true	178	86	25	2	23	0.08	2	0	2	0
rico	65	null	null	null	true	178	42	14	2	12	0.142857	2	0	2	0
rico	66	null	null	null	true	65	40	6	2	4	0.333333	1	1	0	1
rico	67	null	null	null	true	68	24	8	5	3	0.625	5	3	2	0.6
rico	68	null	null	null	true	19	16	4	4	0	1	2	2	0	1
rico	69	null	null	null	true	23	16	4	4	0	1	1	1	0	1
rico	70	null	null	null	true	60	32	9	3	6	0.333333	1	1	0	1
rico	71	null	null	null	true	58	11	0	0	0	null	0	0	0	null
rico	72	null	null	null	true	136	19	4	2	2	0.5	2	2	0	1
rico	73	null	null	null	true	41	20	5	4	1	0.8	3	3	0	1
rico	74	null	null	null	true	70	38	11	3	8	0.272727	1	1	0	1
rico	75	null	null	null	true	57	36	7	3	4	0.428571	1	1	0	1
rico	76	null	null	null	true	32	24	8	4	4	0.5	4	1	3	0.25
rico	77	null	null	null	true	43	30	1	1	0	1	1	1	0	1
rico	78	null	null	null	true	71	49	13	4	9	0.307692	2	2	0	1
rico	79	null	null	null	true	57	38	6	2	4	0.333333	1	1	0	1
rico	80	null	null	null	true	59	31	8	1	7	0.125	1	1	0	1
rico	81	null	null	null	true	32	23	7	4	3	0.571429	3	1	2	0.333333
rico	82	null	null	null	true	58	22	4	2	2	0.5	3	2	1	0.666667
rico	83	null	null	null	true	43	33	0	0	0	null	0	0	0	null
rico	84	null	null	null	true	63	37	7	0	7	0	0	0	0	null
rico	85	null	null	null	true	66	41	11	3	8	0.272727	9	1	8	0.111111
rico	86	null	null	null	true	21	15	1	0	1	0	0	0	0	null
rico	87	null	null	null	true	58	17	4	2	2	0.5	3	2	1	0.666667
rico	88	null	null	null	true	24	22	4	2	2	0.5	2	1	1	0.5
rico	89	null	null	null	true	8	6	0	0	0	null	0	0	0	null
rico	92	null	null	null	true	20	14	3	0	3	0	0	0	0	null
rico	94	null	null	null	true	47	31	9	7	2	0.777778	5	5	0	1
rico	95	null	null	null	true	60	56	18	2	16	0.111111	1	1	0	1
rico	96	null	null	null	true	69	55	10	2	8	0.2	1	1	0	1
rico	97	null	null	null	true	12	10	0	0	0	null	0	0	0	null
rico	98	null	null	null	true	39	36	5	2	3	0.4	1	1	0	1
rico	99	null	null	null	true	19	16	2	2	0	1	1	1	0	1
rico	100	null	null	null	true	69	31	4	2	2	0.5	1	1	0	1
rico	101	null	null	null	true	35	24	2	0	2	0	0	0	0	null
rico	102	null	null	null	true	29	25	4	2	2	0.5	1	1	0	1

End of preview. Expand in Data Studio

a11ydebt per-screen statistics

How many of the elements on each screen of two GUI-agent corpora need a screen-reader label, and how many have one. One row per screen, for every screen of Rico (2017, 66,261 screens) and every step of AndroidControl (2024, 99,131 screens in 20 shards). The rows hold identifiers and counts only: no screenshots, no view hierarchies, no UI text, no class names and no app names.

The counts apply the operational codebook of the IEEE AISSA 2026 paper Trained on What Screen Readers Cannot Read: Accessibility Debt in the GUI Corpora Behind Computer-Use Agents, as implemented in the a11ydebt package. The a11ydebt Space is built on this data.

from datasets import load_dataset

rico = load_dataset("obadakr/a11ydebt-stats", split="rico")
androidcontrol = load_dataset("obadakr/a11ydebt-stats", split="androidcontrol")

Fields

Column	Type	Meaning
`dataset`	string	`rico` or `androidcontrol`
`screen_id`	string	Rico: the UI number (`combined/<screen_id>.json`). AndroidControl: `<episode_id>_<step_index>`
`episode_id`	int64	AndroidControl episode id; empty for Rico
`step_index`	int32	AndroidControl step within the episode (index into `accessibility_trees`); empty for Rico
`shard`	int16	AndroidControl shard number, 0 to 19 (`android_control-000NN-of-00020`); empty for Rico
`in_paper_sample`	bool	In the paper's sample: every Rico screen, and AndroidControl shards 0 to 9
`n_nodes`	int32	Nodes in the screen's tree
`n_included`	int32	Nodes that count: visible to the user and with positive area
`needs_label_candidates`	int32	Included elements that need a label: images, and interactive elements with no visible text (containers and web views are exempt)
`needs_label_labelled`	int32	Of those, the ones with a non-empty content description
`needs_label_unlabelled`	int32	Candidates minus labelled: the screen's accessibility debt
`needs_label_coverage`	float64	Labelled / candidates; empty when there are no candidates
`icon_like_candidates`	int32	Included elements that are a tappable image or button with no visible text (the paper's strictest measure)
`icon_like_labelled`	int32	Of those, the ones with a content description
`icon_like_unlabelled`	int32	Candidates minus labelled
`icon_like_coverage`	float64	Labelled / candidates; empty when there are none

needs_label is the paper's primary measure and icon_like its strictest. a11ydebt also counts an interactive_image population, which is not in the paper and is left out here.

Corpus totals

These sums of the rows equal the paper's Table 1 for Rico and AndroidControl's 10 shards.

Split, rows	Elements counted	Needs a label	Labelled	Coverage	Icon-like	Icon-like coverage
`rico`, all 66,261	2,862,639	502,585	89,146	17.74%	127,552	33.20%
`androidcontrol`, the 49,892 with `in_paper_sample`	6,831,063	1,462,544	837,536	57.27%	194,529	62.75%
`androidcontrol`, all 99,131	13,494,134	2,894,185	1,654,820	57.18%	390,697	62.34%

Share of screens with at least one unlabelled element that needs a label: Rico 85.9% (56,941), AndroidControl 99.6% (49,692 in the paper's sample; 98,710 of all 99,131). The app-cluster bootstrap intervals need each screen's app, which this dataset leaves out; they are in a11ydebt's published tables (results/v0.1/coverage.csv in the repository).

How it was computed

Rico. a11ydebt audit rico <rico folder> --workers 8 --out audit_rico, over the unique_uis.tar.gz view hierarchies and ui_details.csv.
AndroidControl. The free Kaggle CPU kernel in kaggle/androidcontrol_parse/, run 1 for shards 00000 to 00009 and run 2 for 00010 to 00019. It verifies each shard's checksum and writes per-screen tables (screens/*.parquet).
This dataset. python space/build_stats.py --rico audit_rico/screens.csv --androidcontrol run1/screens run2/screens --out data. The script reads only the id and count columns, refuses the kernel's node tables, and fails if any string column holds anything but a dataset name or an id.

The runs used a11ydebt 0.1.0 before its release; the codebook and counting code are unchanged in the released 0.1.1, whose aggregate tables are byte-identical to these runs'. Every row was also checked against the package's own per-screen tables, screen by screen. Parsing is deterministic, so a rerun gives the same rows.

Provenance. Every value is computed by code from the source corpora. No language model and no human annotation produced or edited any row.

Looking a screen up

The ids let you find a screen in your own copy of the corpus, which you obtain under its terms:

Rico: combined/<screen_id>.json (and <screen_id>.jpg) in unique_uis.tar.gz.
AndroidControl: open shard android_control-000NN-of-00020 (NN is shard), find the record whose episode_id matches, and read accessibility_trees[step_index].

Limits

Content descriptions only. Coverage counts a non-empty content-desc. It does not count other ways Android can name an element, such as hint text or a label from a nearby view, and it does not judge whether a label is meaningful.
Roles come from the class name alone, as in the paper.
Small denominators. A screen's coverage over one or two candidates is a coarse number; filter on the candidate counts when you compare screens.
Snapshots. Rico is a 2017 crawl of Android apps and AndroidControl was released in 2024. Neither is a random sample of the apps people use today.
Not a full accessibility audit. Contrast, touch-target size and focus order are not measured.

Licences

These statistics: CC BY 4.0. Please credit the paper and the software (below), and the two source corpora.
Rico (Deka et al., UIST 2017) is distributed by the University of Illinois under its own research-use terms (https://www.interactionmining.org/archive/rico/copyright.txt), which are not an open licence. This dataset contains none of Rico's files, screenshots, hierarchies or text; Rico's terms still govern the underlying data, which you must obtain from its source.
AndroidControl (Li et al., NeurIPS 2024 Datasets and Benchmarks) is Apache-2.0, Copyright 2024 The Google Research Authors.
The code that produced the rows (a11ydebt) is Apache-2.0.

Citation

@inproceedings{kraishan2026trained,
  title     = {Trained on What Screen Readers Cannot Read: Accessibility Debt in the {GUI}
               Corpora Behind Computer-Use Agents},
  author    = {Kraishan, Obada and Jitkajornwanich, Kulsawasd},
  booktitle = {IEEE AISSA 2026},
  year      = {2026},
  note      = {To appear}
}

@software{kraishan2026a11ydebt,
  author  = {Kraishan, Obada},
  title   = {a11ydebt: Audit Accessibility Debt in GUI-Agent Corpora},
  year    = {2026},
  url     = {https://github.com/obadaKraishan/a11ydebt},
  doi     = {10.5281/zenodo.24121869}
}

Please also cite the source corpora:

Deka, B., Huang, Z., Franzen, C., Hibschman, J., Afergan, D., Li, Y., Nichols, J., and Kumar, R. (2017). Rico: A Mobile App Dataset for Building Data-Driven Design Applications. UIST '17. https://doi.org/10.1145/3126594.3126651
Li, W., Bishop, W., Li, A., Rawles, C., Campbell-Ajala, F., Tyamagundlu, D., and Riva, O. (2024). On the Effects of Data Scale on UI Control Agents. NeurIPS 2024 Datasets and Benchmarks. arXiv:2406.03679

Downloads last month: 34

Space using obadakr/a11ydebt-stats 1

Paper for obadakr/a11ydebt-stats

On the Effects of Data Scale on Computer Control Agents

Paper • 2406.03679 • Published Jun 6, 2024