This dataset consists of screenshots of third-party websites, mobile applications, desktop environments, and frames extracted from third-party video recordings. Access is governed by the Restricted Research-Only License and Terms of Access (the LICENSE, full text: https://huggingface.co/datasets/PrentisAI/ScreenRef-Annotations/blob/main/LICENSE). If anything below differs from the LICENSE, the LICENSE prevails.

By requesting access you agree to the LICENSE, and in particular that:

you will use the dataset only for Academic Research as defined in the LICENSE: research by an individual affiliated with a university, college, or public or non-profit research institution, intended for publication, and not carried out for or on behalf of any commercial entity;
you will not redistribute the dataset or any derived copy, except as the LICENSE allows (for example, storage on your institution's infrastructure for approved users only, or up to ten images in a publication with personal information redacted);
you will not attempt to identify, locate, contact or profile any person whose personal information appears in the data;
you understand that screenshots were captured automatically and may incidentally show personal information; you will not disseminate it, and you are asked to report it to requests@prentis.ai;
any model you release that was trained on the dataset will be released only under terms that prohibit commercial use, and will cite the dataset;
when notified, you will delete removed material, or all copies on termination, within 30 days;
the dataset is provided "as is", without warranty, and the LICENSE is governed by the laws of the State of California.

ScreenRef

GUI grounding instructions paired with screenshots from desktop, web and mobile environments, plus an abstention subset in which the referenced UI element is not present and the correct behaviour is to decline to act.

Access to this dataset is gated and granted manually. Use is governed by the LICENSE (Restricted Research-Only License and Terms of Access): Academic Research only, no redistribution, no re-identification. Please read Data provenance before requesting.

The annotations, task recipes, per-image checksums and all scripts are also available without access approval in PrentisAI/ScreenRef-Annotations (no images, 484 MB), so you can inspect the data before requesting access.

Intended use

ScreenRef is intended for Academic Research, as defined in the LICENSE, on GUI agents, in particular:

training and fine-tuning models that locate a user-interface element on a desktop, web or mobile screenshot from a natural-language instruction (GUI grounding);
studying abstention, i.e. declining to act when the referenced element is not on screen;
analysing grounding behaviour across platforms, screen resolutions and instruction types.

Out of scope: any commercial use; identifying, profiling or tracking people or accounts that appear in the screenshots; and tasks beyond single-step grounding, such as multi-step planning, typing or scrolling, which the dataset does not contain.

Composition

Subset	Instructions	Unique screenshots	Notes
`Crawl_*` (web / Android / Ubuntu)	284,531	237,143	screenshots captured by our own pipeline
`traj_*` (video-derived frames)	326,363	326,363	1 instruction per frame
`*_refusal` (abstention)	22,475	22,475	correct answer is `abstain`
Total	633,369	585,981	447 shards, ~327 GiB

Screenshot totals by source tag:

Tag	Screenshots	Tag	Screenshots
`Crawl_Ubuntu`	100,761	`traj_parallel`	170,983
`Crawl_1M_filter`	92,919	`traj_mr`	76,799
`Crawl_Android_part1`	43,463	`traj_hr`	65,858
`refusal_parallel`	16,872	`traj_endfix`	12,723
`refusal_hr`	4,095	`refusal_endfix`	1,508

58 task groups in total. Because some screenshots carry several instructions, there are 8.1% more rows than unique images (all re-use is within Crawl_*; traj_* and the refusal subset are strictly 1 : 1).

Schema

Column	Type	Description
`image`	`Image()`	the screenshot
`width`, `height`	`int32`	native resolution of the screenshot
`conversations`	`string`	JSON list of `{"from": "human"/"gpt", "value": ...}`
`task`	`string`	task group the row belongs to
`conv_style`	`string`	conversation template identifier

conversations is stored as a raw JSON string and is not re-interpreted by the loader. Coordinates are integers normalized to a 1000 × 1000 space, not pixels — for tool-call coordinate and <box> alike. To get pixels: x_px = x / 1000 * width, y_px = y / 1000 * height.

Action space

Answers are <tool_call> blocks. The tool name follows the platform and the action distinguishes grounding from abstention:

Platform	Tool	Grounding action	Abstention action
desktop	`computer_use`	`left_click` + `coordinate`	`abstain`
web	`browser_use`	`left_click` + `coordinate`	`abstain`
mobile	`mobile_use`	`click` + `coordinate`	`abstain`

Row counts: left_click 541,793 · click 58,782 · abstain 22,475. The three *_text_box tasks (10,319 rows) answer with <ref>…</ref><box>[[x1,y1,x2,y2]]</box> instead of a tool call. There are no typing, scrolling, dragging or keyboard actions.

Example of an abstention row:

{"from": "human", "value": "<image>\nLocate and click the \"Window\" menu item."}
{"from": "gpt",   "value": "<tool_call>\n{\"name\": \"computer_use\", \"arguments\": {\"action\": \"abstain\"}}\n</tool_call>"}

Usage

from datasets import load_dataset

ds = load_dataset("PrentisAI/ScreenRef", split="train", streaming=True)  # requires access + token
row = next(iter(ds))
row["image"]          # PIL image
row["task"]           # e.g. "traj_desktop_refusal__hr"

Every shard holds rows from exactly one task, and shard_index.tsv lists the task, row count and size of each shard. To work with a subset, download only its shards instead of filtering the full 327 GiB. For example, abstention rows only:

import csv
from datasets import load_dataset

# get shard_index.tsv first, e.g. hf_hub_download("PrentisAI/ScreenRef", "shard_index.tsv", repo_type="dataset")
rows = csv.DictReader(open("shard_index.tsv"), delimiter="\t")
shards = [r["shard"] for r in rows if "refusal" in r["task"]]   # 12 shards, 5.8 GiB

refusal = load_dataset("PrentisAI/ScreenRef", data_files=shards, split="train")  # 22,475 rows

Data provenance

The two halves of the dataset were obtained differently, but in both cases the content shown on screen belongs to third parties, not to the maintainers. Specifically:

Crawl_* — screenshots collected by the maintainers with their own automated pipeline: publicly reachable web pages rendered in a browser, Android applications running in an emulator (1080×2400) and Ubuntu desktop applications running in a virtual machine (1920×1080). They are not copied from any existing public dataset. Crawl_1M_filter covers a long tail of websites (577 distinct domains appear in the annotation text alone, 72% of them only once, spanning .com/.io/.org/.gov/.de/.ru/.dev/.edu and others).
traj_* and the refusal subset — frames extracted from publicly available YouTube tutorial videos of software being operated. The directory name in each image path is the source video's YouTube ID.

Rights in the depicted websites, applications, user interfaces and video recordings remain with their respective owners. The maintainers make no claim of ownership over that content. Prentis AI owns only the annotations and the compilation; the LICENSE licenses those and sets the terms of access for the screenshots.

Takedown / rights requests: if you are a rights holder and wish content removed, contact the maintainers (see Contact). Removal requests are actioned and a corrected revision published; access may be suspended meanwhile.

Because screenshots were captured automatically, some may incidentally show personal information. Please report any you encounter to the address under Contact.

Known limitations

Instruction verb and action can disagree. Every grounding answer is a single click, but about 3.6% of the video-derived instructions name a different gesture — right-click (5,960), double-click (3,659), hover (1,650), drag (277) or long-press (223). Treat these as location supervision only; the dataset does not teach gesture selection.
Coordinates are normalized to 0–1000 on both axes, while resolutions are heterogeneous (1080×2400, 1920×1080, 2560×1440, 3840×2160 and others); convert with width / height if you need pixels.

Licence

See the LICENSE (the same file is included here and, publicly, in the annotation repository). In short: Academic Research only (as defined there), no commercial use, no redistribution beyond the listed exceptions, no re-identification, and deletion within 30 days of notice. Models trained on the dataset may be released under non-commercial terms. The LICENSE grants no rights in the depicted software, web content or recorded material.

Contact

Access review, personal-information reports and takedown requests: requests@prentis.ai

Citation

The accompanying paper is currently under review. Citation information will be added upon publication.

Downloads last month: 32

Total file size:

351 GB