PrentisAI/ScreenRef-Annotations

ScreenRef — Annotations

Annotations only. This repository contains no screenshots. It holds every annotation of ScreenRef (484 MB); the screenshots are in the gated repository PrentisAI/ScreenRef. The annotations are licensed under the same LICENSE as the screenshots: Academic Research only, no redistribution.

Contents

Path	What it is	Size
anno/	58 JSONL files — all tasks (the full recipe), 633,369 rows	240 MB
anno_point/	46 JSONL files — same screens, point-format targets, 610,894 rows over 49 tasks	216 MB
meta/	Task recipes naming each task's annotation file, image root and conversation style: meta_full.json (all 58 tasks), meta_non_refusal.json (49) and meta_refusal.json (9)	34 KB
scripts/	The Parquet exporter and unpacker	11 KB
stats/task_table_*.tsv	Per-task rows, unique images, conversation style, repeat weight	15 KB
stats/shard_index.tsv	Which task each of the 447 Parquet shards holds, with rows and bytes	34 KB
stats/image_manifest.tsv.gz	All 585,981 screenshots: relative path, byte size, SHA-256	28 MB
stats/dataset_stats.json	Machine-readable summary statistics	1 KB

Composition (full recipe)

Group	Tasks	Rows	Image roots
Trajectory frames (traj_*)	20	326,363	traj_endfix, traj_hr, traj_mr, traj_parallel
Dense crawled screens (v3c10_*)	27	282,992	Crawl_1M_filter, Crawl_Android_part1, Crawl_Ubuntu
Refusal / abstention	9	22,475	refusal_endfix, refusal_hr, refusal_parallel
Spatial relations	2	1,539	Crawl_Android_part1, Crawl_Ubuntu
Total	58	633,369	585,981 distinct screenshots

The point recipe re-expresses the 49 non-refusal tasks with <point> targets; its three *_text_box tasks reuse the box-format files in anno/.

Record schema

Every row has exactly four fields:

{
  "image": "ubuntu/76a3a3711e8c69518ca05db6e516592a2a8d0bd08fe234cd26644ae9d3a3ab0f.png",
  "conversations": [
    {"from": "human", "value": "<image>\nMake your selection: `Recent`."},
    {"from": "gpt",   "value": "<tool_call>\n{\"name\": \"computer_use\", \"arguments\": {\"action\": \"left_click\", \"coordinate\": [86, 244]}}\n</tool_call>"}
  ],
  "width": 1920,
  "height": 1080
}

• image is relative to the root given in the task's meta entry.
• width / height are the screenshot's pixel dimensions. Coordinates are integers normalized to a 1000 × 1000 space, not pixels: x_px = x / 1000 * width, y_px = y / 1000 * height.
• Every answer is a single click or an abstention; the three *_text_box tasks answer with <ref>…</ref><box>[[x1,y1,x2,y2]]</box> instead.

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

Rows: left_click 541,793 · click 58,782 · abstain 22,475 · box 10,319. Abstention rows (only in the nine *_refusal* tasks) pair a screenshot with an instruction whose target is not on screen.

Getting the screenshots

The screenshots are in PrentisAI/ScreenRef (gated, 447 Parquet shards, 327 GiB). To train, load that repository directly — every row already carries its screenshot and conversation.

To use this repository's recipes instead (for example point), unpack the shards into an images/ tree next to anno/, anno_point/ and meta/:

hf download PrentisAI/ScreenRef --repo-type dataset --include "data/*" --local-dir ./screenref-parquet
python scripts/parquet_to_tree.py --shards ./screenref-parquet/data \
    --meta meta/meta_full.json --out . --verify stats/image_manifest.tsv.gz

Rebuilding the Parquet release

With images/ in place, this regenerates the 447 shards of PrentisAI/ScreenRef byte for byte (about 11 minutes on 8 cores):

bash scripts/hf_export/reexport.sh meta/meta_full.json parquet_out/data

Data provenance

• traj_* and refusal_* (349K rows, 55%). Frames from publicly available YouTube screen-recording tutorials; the first directory of each image path is the source video's YouTube ID. We hold no rights in the recorded material.
• Crawl_* (284K rows). Screenshots collected with our own automated pipeline — web pages in a browser, Android apps in an emulator, Ubuntu applications in a virtual machine — not copied from any public dataset. The depicted software and web content belongs to its respective owners.
• Annotations and the compilation are our own work. The LICENSE licenses them and sets the terms of access for the screenshots.

Rights holders who want material removed can write to the address under Contact; we will remove the affected screenshots and rows and reissue. Screenshots may incidentally show personal information; please report any you encounter, and do not attempt to re-identify anyone who appears in the data.

Known limitations

• Instruction verb and action can disagree. About 3.6% of the video-derived instructions name a gesture other than a click — right-click (5,960), double-click (3,659), hover (1,650), drag (277) or long-press (223) — while the answer is always a single click. Treat these as location supervision only.
• Trajectory rows are not trajectories. Each row is an independent single-step example; traj_* refers to where the frames came from.

Licence

See LICENSE (Restricted Research-Only License and Terms of Access). The annotations are licensed for Academic Research only: research by individuals at universities or public or non-profit research institutions, intended for publication and not carried out for a commercial entity. You may not redistribute them, but you may quote a few records in a publication and release new annotations of your own that refer to records by identifier. The screenshots are available only to approved users of PrentisAI/ScreenRef. The LICENSE grants no rights in the depicted software, web content or recorded material.

Citation

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

Contact

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