cua-lite/OS-Atlas

---
license: other
tags:
- cua-lite
- gui
- sft
task_categories:
- image-text-to-text
configs:
- config_name: default
data_files:
- split: train
path:
- "*/*/[t]rain.parquet"
- "*/*/train/*.parquet"
- "*/*/train/*/*.parquet"
- split: validation
path:
- "*/*/[v]alidation.parquet"
- "*/*/validation/*.parquet"
- "*/*/validation/*/*.parquet"
- config_name: desktop
data_files:
- split: train
path:
- "desktop/*/[t]rain.parquet"
- "desktop/*/train/*.parquet"
- "desktop/*/train/*/*.parquet"
- split: validation
path:
- "desktop/*/[v]alidation.parquet"
- "desktop/*/validation/*.parquet"
- "desktop/*/validation/*/*.parquet"
- config_name: mobile
data_files:
- split: train
path:
- "mobile/*/[t]rain.parquet"
- "mobile/*/train/*.parquet"
- "mobile/*/train/*/*.parquet"
- split: validation
path:
- "mobile/*/[v]alidation.parquet"
- "mobile/*/validation/*.parquet"
- "mobile/*/validation/*/*.parquet"
- config_name: web
data_files:
- split: train
path:
- "web/*/[t]rain.parquet"
- "web/*/train/*.parquet"
- "web/*/train/*/*.parquet"
- split: validation
path:
- "web/*/[v]alidation.parquet"
- "web/*/validation/*.parquet"
- "web/*/validation/*/*.parquet"
- config_name: desktop-grounding-bbox
data_files:
- split: train
path:
- "desktop/grounding-bbox/[t]rain.parquet"
- "desktop/grounding-bbox/train/*.parquet"
- "desktop/grounding-bbox/train/*/*.parquet"
- split: validation
path:
- "desktop/grounding-bbox/[v]alidation.parquet"
- "desktop/grounding-bbox/validation/*.parquet"
- "desktop/grounding-bbox/validation/*/*.parquet"
- config_name: mobile-grounding-bbox
data_files:
- split: train
path:
- "mobile/grounding-bbox/[t]rain.parquet"
- "mobile/grounding-bbox/train/*.parquet"
- "mobile/grounding-bbox/train/*/*.parquet"
- split: validation
path:
- "mobile/grounding-bbox/[v]alidation.parquet"
- "mobile/grounding-bbox/validation/*.parquet"
- "mobile/grounding-bbox/validation/*/*.parquet"
- config_name: web-grounding-bbox
data_files:
- split: train
path:
- "web/grounding-bbox/[t]rain.parquet"
- "web/grounding-bbox/train/*.parquet"
- "web/grounding-bbox/train/*/*.parquet"
- split: validation
path:
- "web/grounding-bbox/[v]alidation.parquet"
- "web/grounding-bbox/validation/*.parquet"
- "web/grounding-bbox/validation/*/*.parquet"
---

# cua-lite/OS-Atlas

cua-lite preprocessed version of OS-Atlas (OS-Copilot/OS-Atlas-data). grounding:bbox across three platforms and five sub-sources: desktop (windows, linux, macos), mobile (amex), web (fineweb).

## Origin

- [https://huggingface.co/datasets/OS-Copilot/OS-Atlas-data](https://huggingface.co/datasets/OS-Copilot/OS-Atlas-data)

## Load via `datasets`

```python
from datasets import load_dataset

# entire dataset
ds = load_dataset("cua-lite/OS-Atlas")

# just one platform
ds = load_dataset("cua-lite/OS-Atlas", "desktop")

# just one (platform, task_type) cohort
ds = load_dataset("cua-lite/OS-Atlas", "desktop-grounding-bbox")
```

You can also filter by `metadata.platform` / `metadata.task_type` /
`metadata.others.*` after loading; every row carries a rich `metadata`
struct (see schema below).

## Schema

Each row has these columns:

| column | type | notes |
|---|---|---|
| `image_ids` | list[string] | content-addressed ids (`<sha256>.<ext>`), enables cross-parquet / cross-dataset dedup |
| `images` | list[Image] | bytes embedded at HF push time; matches `image_ids` index-for-index |
| `messages` | list[struct] | OpenAI-style turns with `role` + structured `content` |
| `metadata` | struct | `{platform, task_type, split, others{...}}` |

Coordinate values in `messages` are normalized to `[0, 1000]` integers.

## Layout

```
<platform>/<task_type>/<split>.parquet # single-variant cohort
<platform>/<task_type>/<split>/<variant>.parquet # multi-variant cohort
<platform>/<task_type>/<split>/shard-NNNNN-of-NNNNN.parquet # + sharded single-variant
<platform>/<task_type>/<split>/<variant>/shard-NNNNN-of-NNNNN.parquet # + sharded multi-variant
```

- `platform` ∈ {desktop, mobile, web}
- `task_type` directory uses a hyphen where the metadata value uses a colon: `grounding-action/` → `grounding:action`
- `split` ∈ {train, validation} — `validation` is an in-distribution held-out slice (never used in training); `test` is reserved for out-of-distribution benchmark datasets

## Stats

| platform | task_type | variant | train | validation |
|---|---|---|---:|---:|
| desktop | grounding:bbox | linux | 42,327 | 817 |
| desktop | grounding:bbox | macos | 17,958 | 440 |
| desktop | grounding:bbox | windows | 1,073,175 | 2,000 |
| mobile | grounding:bbox | amex | 1,200,434 | 2,000 |
| mobile | grounding:bbox | aw | 88,078 | 1,753 |
| mobile | grounding:bbox | ricosca | 169,858 | 3,417 |
| mobile | grounding:bbox | uibert | 16,353 | 307 |
| mobile | grounding:bbox | widget | 99,425 | 2,000 |
| web | grounding:bbox | fineweb | 6,639,491 | 2,000 |
| web | grounding:bbox | seeclick | 2,112,523 | 2,000 |

## Image storage

Images are content-addressed by SHA-256 and deduplicated within this repo.
The `images` column on HuggingFace embeds raw bytes so the Hub viewer
renders thumbnails and `datasets.load_dataset` works out of the box.

For local workflows (SFT export, cross-dataset dedup, split rebalancing),
run [`reverse.py`](https://github.com/cua-lite/cua-lite/tree/main/scripts/hf_upload)
on a cloned repo: it extracts each unique `image_id` once to a shared
`image_store/<hash[:2]>/<hash>.<ext>` and rewrites the parquets to drop
the `images` column, so rows reference images by hash id only. The shared
store is reusable across datasets — the same image in two repos lands in
one file.

- Total unique images: **1,312,118**
- Store size: **496.53 GB**

## Notes

Images are heavily reused: each screenshot is typically referenced by dozens of bbox labels. Content-addressed storage collapses this to a far smaller unique-image count than the ~3.58M row count suggests.

## License & citation

See original dataset (OS-Copilot/OS-Atlas-data)

See https://huggingface.co/datasets/OS-Copilot/OS-Atlas-data