Datasets:
The dataset viewer is not available for this subset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 89, in _split_generators
pa.Table.from_pylist(cast_to_python_objects([example], only_1d_for_numpy=True))
~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "pyarrow/table.pxi", line 2044, in pyarrow.lib._Tabular.from_pylist
File "pyarrow/table.pxi", line 6401, in pyarrow.lib._from_pylist
return cls.from_arrays(arrays, names, metadata=metadata)
File "pyarrow/table.pxi", line 4868, in pyarrow.lib.Table.from_arrays
converted_arrays = _sanitize_arrays(arrays, names, schema, metadata,
File "pyarrow/table.pxi", line 1611, in pyarrow.lib._sanitize_arrays
converted_arrays = _schema_from_arrays(arrays, names, metadata,
File "pyarrow/table.pxi", line 1592, in pyarrow.lib._schema_from_arrays
val = array(val)
File "pyarrow/array.pxi", line 400, in pyarrow.lib.array
File "pyarrow/array.pxi", line 46, in pyarrow.lib._sequence_to_array
chunked = GetResultValue(
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
return check_status(status)
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowInvalid: cannot mix list and non-list, non-null values
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Glance-GUI
GUI grounding data: screenshots of desktop applications and web pages, with requests that are either present (with the element's box) or absent. The current release is guisyn-v0, published before as Vision-Decision/guisyn-v0-clean; its original card is in provenance/. Image-grounding data on COCO is in Glance-COCO.
This is synthetic GUI-grounding data for a "presence gate plus box" model. Each screenshot comes with many requests. A request is either present, with the element's box, or absent; absent requests are hard negatives that were verified against the full accessibility tree.
This is the scrubbed public version of the internal guisyn-v0 (19,293 screenshots). It holds 18,535 screenshots (12,730 desktop,
5,805 web) and 1,080,602 requests.
Sources
- desktop. 19 open-source Linux applications run under Xvfb: LibreOffice Writer, Calc, Impress and Draw; GIMP;
Inkscape; Krita; FreeCAD; KiCad pcbnew and eeschema; VSCodium; Firefox; Thunderbird; Kdenlive; Shotcut;
Audacity; darktable; QGIS; Wireshark.
- Resolutions are 1920×1080, 2560×1440, 2560×1600 and 3840×2160 (including 2× UI scale).
- The applications were driven by random but safe interactions on synthetic documents.
- Element boxes, names and roles come from the AT-SPI accessibility tree, filtered for visibility, occlusion, clipping, hit-testing, pixel content and duplicates.
- web. Playwright Chromium screenshots of public Tranco top sites (list V38GN) at DPR 1 and 2. The crawler respected robots.txt and did no form interaction. Elements come from the visible-viewport DOM.
- Negatives.
negative_harddescribes an element seen in another state of the same app, or another viewport of the same site, and verified absent from this screenshot's full accessibility tree or visible DOM (exact, fuzzy with rapidfuzz ≥ 80, and substring name match).negative_easycomes from other apps.
- No ScreenSpot-v2, ScreenSpot-Pro or VenusBench-GD image is included (dHash and thumbnail check).
Layout
shards/{desktop,web}/<stem>.jsonl: one record per screenshot.<stem>.images.tar:images/{source}/{id}.webp, lossless WebP.<stem>.a11y.tar:{id}.json.gz; the accessibility tree and X stacking order (desktop), or DOM elements and visible texts (web).
Record: {image, W, H, source, app, url, questions: [{text, present, box [x1,y1,x2,y2] px | null, kind: name|function|position|icon|negative_hard|negative_easy, meta}], uncaptioned: [{box, role}], a11y, meta}.
stats.json has the counts per app and per kind.
What was removed, and why
| removed | how much | why |
|---|---|---|
screenshots whose OCR text (Tesseract 5, --psm 11) matched an e-mail, phone, key or high-entropy pattern |
229 | rendered PII or secrets cannot be redacted from pixels reliably, so the whole screenshot was dropped, never inpainted |
| screenshots whose OCR or tree text contains a public IPv4 address, plus every page from IP-lookup, VPN and geolocation sites | 529 | some sites display the crawler's own public IP address; sample network captures (Wireshark) and documents show public addresses |
| raw page HTML in the web a11y files | 5,805 fields | it carried third-party sites' embedded keys (Algolia, IPInfo, Disqus, Storyblok, LocationIQ, New Relic) |
| requests whose text or element name matched a secret or PII pattern | 453 | e.g. a link labelled with an e-mail address |
other matching strings in a11y and record text, replaced by [REDACTED] |
8,834 | e-mail addresses, phone numbers, key-like and high-entropy tokens, home paths |
internal screenshot ids, replaced by short ids <app>-<n> |
all | the original ids embedded job seeds and timestamps |
raw_jobs/ (raw job logs), paraphrases_deepseek.json |
— | logs; a paraphrase cache keyed by raw element names, which contained an e-mail address |
Licence and provenance
| part | licence |
|---|---|
| requests, boxes, labels, a11y-derived annotations (ours) | CC BY 4.0 |
| desktop screenshots | screenshots of open-source applications on synthetic content; the applications' own licences apply to their UI artwork |
| web screenshots and web element text | rendered third-party content: copyright and site terms remain with the site owners; no licence is granted here. Provided for non-commercial research on GUI grounding only. Site owners can request removal through the repository's community tab |
Scan method
Before publication, every text file (records, and every a11y JSON before packing) went through two passes of secret and PII scanning, and both came back with 0 findings:
- Fixed patterns.
sk-,hf_,AIza,token=/api_key=, Google access tokens (ya29) and Bearer tokens, GitHub, AWS, Slack and Stripe keys, JWTs, private-key headers, and provider keys for Algolia, IPInfo, Disqus, Storyblok, LocationIQ and New Relic. - Generic patterns. High-entropy strings (32 or more characters, Shannon entropy above 4.5, pure hex hashes excluded), e-mail addresses, phone numbers and home-directory paths.
In addition, every screenshot was OCR'd, and the release was checked for the crawling host's own IP addresses. The scanner was validated on a positive control with planted keys.
Intended use
Training and evaluating GUI grounding with abstention: given a screenshot and a request, return the element's box or say it is not there.
Limitations
- OCR misses small or low-contrast text, so a dropped-screenshot filter is a strong reduction, not a guarantee. Report anything that slipped through, and it will be removed.
- Redaction can leave gaps such as
[REDACTED]in element names. - Element names come from accessibility metadata, which is sometimes missing or wrong (for example, icons without
names are listed under
uncaptioned, not used as requests). - Desktop states come from random interaction. Some apps (Thunderbird, Audacity) lost more screenshots to the filter than others.
Citation
@misc{guisyn_v0,
title = {guisyn-v0: synthetic GUI grounding with verified hard negatives},
author = {Vision-Decision},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/Vision-Decision/Glance-GUI}}
}
- Downloads last month
- -