README.md

---
license: cc-by-nc-4.0
pretty_name: CaptchaArena
task_categories:
- image-text-to-text
language:
- en
tags:
- captcha
- gui-agent
- multimodal
- vision-language
- benchmark
size_categories:
- 10K<n<100K
extra_gated_heading: "Access to CaptchaArena (opens after our arXiv release)"
extra_gated_description: "CaptchaArena is released for non-commercial academic research only (CC-BY-NC-4.0); commercial use is prohibited. Data sharing has not started yet — we will begin granting access once our paper is available on arXiv. You are welcome to submit a request now; requests will be reviewed after the paper is released."
extra_gated_prompt: "⏳ Access is not open yet. We will start sharing CaptchaArena once our paper is posted on arXiv — access requests will be reviewed at that time, so please check back after the paper release. By requesting access you agree to use the dataset solely for non-commercial academic research; any commercial use is prohibited. Requests are reviewed by the dataset authors."
extra_gated_fields:
  Full name: text
  Affiliation / Institution: text
  Email: text
  Intended research use: text
  I confirm I will use this dataset for non-commercial academic research only: checkbox
  I confirm I will NOT use this dataset for any commercial purpose: checkbox
extra_gated_button_content: "Request access"
---

# CaptchaArena

**CaptchaArena** is a multimodal **benchmark** for evaluating GUI / computer-use agents on their ability to *solve* CAPTCHA puzzles. It spans **20 CAPTCHA task families** (point-and-click, counting, rotation alignment, drag-to-fit, press-and-hold, connect-the-icons, …), each rendered as one or more screenshot images with an instruction and a **machine-checkable ground-truth answer**, so agent solutions can be scored automatically. Held-out **Val / Test** splits give a standardized evaluation setting, with a larger **Train** split also included.

> **⚠️ Access & use:** This dataset is **gated** and released for **non-commercial academic research only — commercial use is prohibited** (CC-BY-NC-4.0). Request access above and agree to the terms.

**🔗 Code:** [github.com/X0X0X00/CaptchaArena](https://github.com/X0X0X00/CaptchaArena) — puzzle generation, mock-solving, and evaluation.

## Task families (20)

| | | | |
|---|---|---|---|
| Bingo | Click_Order | Connect_icon | Coordinates |
| Dart_Count | Dice_Count | Geometry_Click | Hold_Button |
| Image_Matching | Image_Recognition | Misleading_Click | Object_Match |
| Patch_Select | Path_Finder | Pick_Area | Place_Dot |
| Rotation_Match | Select_Animal | Slide_Puzzle | Unusual_Detection |

Each family is a distinct CAPTCHA style: some are single-image / single-step (e.g. `Bingo`, `Geometry_Click`), others are multi-image / multi-step (e.g. `Patch_Select`, `Connect_Icon`, `Slide_Puzzle`).

## Counts

| | Train | Val | Test | Total |
|---|---|---|---|---|
| Puzzles | 42,000 | 4,000 | 4,000 | **50,000** |
| Image files (PNG) | 67,332 | 9,935 | 9,867 | **87,134** |

- **2,100** puzzles per task in Train, **200** per task in Val and Test, across all 20 families.
- Image files outnumber puzzles because multi-image families (e.g. `Patch_Select`, `Coordinates`, `Connect_Icon`) render several PNGs per puzzle (options / references / grid tiles).
- Total size on disk ≈ **53.9 GB**.

## Splits & structure

```
CaptchaArena/
├── Train/          # 2100 puzzles per task
│   ├── Bingo_2100/
│   ├── Click_Order_2100/
│   ├── Connect_icon_2100/      # packaged as a .tar (see Notes)
│   ├── …
│   └── Hold_Button_2100/
├── Val/            # 200 puzzles per task
│   ├── Bingo_200/
│   └── …
└── Test/           # 200 puzzles per task
    ├── Bingo_200/
    └── …
```

- **Train** 2,100 puzzles / task; **Val** and **Test** 200 puzzles / task, across all 20 families.
- Folder naming: `<Split>/<Task>_<N>/` where `N` is the target puzzle count.
- Puzzle images are **PNG**. Layout varies by task: single-image families store flat PNGs (e.g. `Train/Bingo_2100/bingo1.png`); multi-image families group each puzzle in its own sub-folder.
- Every task folder also ships two label files — `ground_truth.json` and `ground_truth_cu.json` (see **Ground truth**).

## Ground truth

Every task folder ships its labels as two JSON files, both keyed by image filename:

- **`ground_truth.json`** — base answer + metadata.
- **`ground_truth_cu.json`** — the same entries plus `answer_cu`, a pixel-space **Computer-Use action sequence** (click / drag / …) that solves the puzzle, tagged with `answer_cu_kind` (e.g. `"tool_calls"`).

The answer key(s) are **family-specific**. A few representative shapes:

```jsonc
// grid pick (Select_Animal, Patch_Select, Image_Recognition, …)
"image1.png": {
  "prompt": "Pick a whale",
  "target_object": "whale",
  "grid_size": [2, 3],
  "correct_patches": [1],                 // index(es) of the correct cell(s)
  "description": "A 2x3 grid of images; pick the named one."
}

// option select (Coordinates, Object_Match, …)
"coord_0140_…__J70.png": {
  "prompt": "Using the arrows, move Jerry to the indicated seat",
  "correct_option_index": 4,              // index into option_images
  "option_images": [".../I72.png", ".../H69.png", "…", ".../J70.png"]
}

// rotation (Rotation_Match)
"puzzle_rotation_232_45.json": {
  "prompt": "Use the arrows to rotate the object to match the reference.",
  "answer": 45, "correct_angle": 45,      // degrees
  "reference_image": "ref_232_45.png", "object_base_image": "232.png"
}

// swap-to-line (Bingo)
"bingo2301.png": {
  "prompt": "Exchange two images to line up identical ones.",
  "answer": [[3, 4]],                     // swap cell 3 <-> cell 4
  "grid_size": [3, 3],
  "solution_line": { "vertical": [1, 4, 7] }
}
```

Other families use analogous keys (counting families store the target count; position families a target point + tolerance). The `_cu` file adds the replayable pixel action sequence:

```jsonc
// ground_truth_cu.json — same entry + a ready-to-replay action sequence
"image1.png": {
  "correct_patches": [1],
  "answer_cu": [
    { "action": "click", "arguments": { "x": 640, "y": 413 } },
    { "action": "click", "arguments": { "x": 640, "y": 885 } }   // e.g. confirm / submit
  ],
  "answer_cu_kind": "tool_calls"
}
```

All coordinates — both in `answer_cu` and in any spatial `ground_truth.json` fields — are **image-natural pixels with the origin at the top-left**, on the benchmark's fixed **1280×1080** viewport (the frame the puzzle image is rendered in), so a stored action stays valid regardless of how the image is later displayed or scaled.

## Notes

- **`Train/Connect_icon_2100/`** is shipped as a single archive, `Connect_icon_2100.tar` (~5.3 GB, LFS). Extract it with:
  ```bash
  tar xf Connect_icon_2100.tar
  ```

## Loading

After your access request is approved, log in and download with the Hub client:

```bash
pip install -U huggingface_hub
hf auth login          # required: this dataset is gated
hf download ZHEN-04/CaptchaArena --repo-type dataset --local-dir CaptchaArena
```

Or grab a single split / task folder:

```python
from huggingface_hub import snapshot_download
snapshot_download(
    "ZHEN-04/CaptchaArena", repo_type="dataset",
    allow_patterns=["Val/Bingo_200/*"], local_dir="CaptchaArena",
)
```

## Relation to CaptchaArena-Trajectories

- **ZHEN-04/CaptchaArena** (this repo) — the **puzzles**: images + instruction + ground-truth answer (base + Computer-Use pixel variant).
- [**ZHEN-04/CaptchaArena-Trajectories**](https://huggingface.co/datasets/ZHEN-04/CaptchaArena-Trajectories) — CoT **Computer-Use agent trajectories** that *solve* these puzzles, in per-turn SFT format.

## License

Released under **CC-BY-NC-4.0** (Creative Commons Attribution–NonCommercial 4.0). **Non-commercial academic research use only — commercial use is prohibited.** Access is gated: you must request access and agree to these terms before downloading. Please attribute when using this dataset.

## Citation

If you use CaptchaArena, please cite this repository and the code at [github.com/X0X0X00/CaptchaArena](https://github.com/X0X0X00/CaptchaArena). *(Formal citation / paper reference: TODO.)*