README.md

---
pretty_name: ScreenRef Annotations
license: other
license_name: restricted-research-only
license_link: LICENSE
task_categories:
- image-text-to-text
language:
- en
tags:
- gui-agent
- gui-grounding
- gui-abstention
- annotations-only
- web
- desktop
- mobile
size_categories:
- 100K<n<1M
configs:
- config_name: full
  data_files:
  - split: train
    path: anno/*.jsonl
- config_name: point
  data_files:
  - split: train
    path:
    - anno_point/*.jsonl
    - anno/v3c10_android_text_box.jsonl
    - anno/v3c10_ubuntu_text_box.jsonl
    - anno/v3c10_web_text_box.jsonl
---

# 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`](https://huggingface.co/datasets/PrentisAI/ScreenRef).
The annotations are licensed under the same [LICENSE](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:

```json
{
  "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`](https://huggingface.co/datasets/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/`:

```bash
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
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](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**