CaptchaArena-Trajectories is released for non-commercial academic research only (CC-BY-NC-4.0); commercial use is prohibited. Access requests are reviewed manually by the dataset authors.

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.

CaptchaArena-Trajectories

A Large-Scale, Fine-Grained Dataset for Training Computer-Use Agents on Interactive CAPTCHAs

Zhenhao Zhang^1,*, Zhaoyu Fan², Haohan Ying³, Jingwen Hu³, Hancen Fan¹,
Junhao Zhou⁴, Zitian Chen¹, Linchao Zhu^2,†
¹Columbia University · ²Zhejiang University
³University of Rochester · ⁴University of Illinois at Urbana-Champaign
^*Project lead · ^†Corresponding author

CaptchaArena-Trajectories is a multimodal chain-of-thought (CoT), Computer-Use agent trajectory dataset for supervised fine-tuning (SFT) of GUI / computer-use agents on the task of solving CAPTCHAs. It is the trajectory (training-data) companion to the puzzle dataset ZHEN-04/CaptchaArena: every trajectory here solves one puzzle drawn from CaptchaArena's 20 task types.

Each trajectory is a full agent rollout — the agent observes a browser screenshot, reasons step by step inside <think>…</think>, then emits a Computer-Use action (click / drag / hold / type), observes the next screenshot, and repeats until the puzzle is solved.

Data generation pipeline

The Computer-Use actions are not model-generated: each puzzle's stored answer is compiled into browser actions and replayed in the live environment, and a trajectory is kept only if the page's own verifier accepts the final state. The step-by-step reasoning is then annotated on top of these verified actions:

Stage	Model / actor	Role
1. Generation	GPT-5.4-mini	Given the screenshot at each step and the correct action (described in task terms, not as a coordinate), writes only the CoT `<think>` reasoning that leads to that action.
2. Verification	Gemini 2.5 Flash × 2	An independent judge from a different model family checks each sample twice, reading image and text together; a sample is admitted only if both checks accept it. It rejects reasoning that contradicts the verified solution (e.g. mistaking one animal/object for another), reveals the answer in hindsight, or hedges instead of committing to the action.
3. Repair	GPT-5.4-mini → GPT-5.5 → author	Rejected samples are regenerated with the judge's feedback. Samples still rejected after 3 attempts are checked by GPT-5.5, which also checks every `Click_Order` and `Connect_Icon` sample; disputed samples go back through the same loop. An author manually reviews, repairs, or discards any remaining cases.

Task types (20)

Interaction mode	Task types
Single-click	Geometry_Click, Misleading_Click, Pick_Area, Place_Dot, Select_Animal
Multi-click	Bingo, Click_Order, Image_Recognition, Patch_Select, Unusual_Detection
Arrow-cycle	Connect_Icon, Coordinates, Dart_Count, Image_Matching, Object_Match, Path_Finder, Rotation_Match
Real-time	Hold_Button, Slide_Puzzle
Text entry	Dice_Count

Counts

	Train	Val	Total
Puzzles (trajectories)	42,000	4,000	46,000
Training samples (`.jsonl` lines)	132,053	12,565	144,618
Step screenshots (PNG)	155,173	14,772	169,945 (≈ 50.9 GB)

2,100 puzzles per task in Train, 200 per task in Val, across all 20 types.
The 4,000 Test puzzles (200 per task) are used only for evaluation and are never annotated, so they are not in this repo; the puzzles themselves are in ZHEN-04/CaptchaArena.
The .jsonl files use a per-turn SFT layout: a T-step trajectory is expanded into T training examples — example k carries the conversation prefix through step k, with the loss applied only on the k-th assistant turn, so the expansion factor varies by type (single-step types ≈ 1×, Patch_Select ≈ 8.5×).

Structure

CaptchaArena-Trajectories/
├── train/
│   ├── bingo_sft_2100_thinking_perturn.jsonl        # per-turn CoT training samples
│   ├── bingo_2100/                                   # step screenshots for this task
│   │   └── Bingo_2100_bingo1/
│   │       └── screenshots/
│   │           ├── screenshot_step_0.png
│   │           └── screenshot_step_1.png …
│   ├── … (20 tasks: <task>_sft_2100_thinking_perturn.jsonl + <task>_2100/)
└── val/
    ├── bingo_sft_200_thinking_perturn.jsonl
    ├── bingo_200/
    └── … (20 tasks)

Each split has 20 .jsonl files (one per task) plus 20 image folders. Image paths inside the .jsonl are relative to the repo root (e.g. train/bingo_2100/Bingo_2100_bingo1/screenshots/screenshot_step_0.png), so the data is usable directly after a full download.

Data format

Every line of a .jsonl is one training example:

{
  "messages": [
    { "role": "system",
      "content": "You are a Computer-Use agent solving exactly one CAPTCHA puzzle on the CaptchaArena benchmark. …" },
    { "role": "user", "content": [
        { "type": "text",  "text": "Here is the current state of the browser. Solve this puzzle:" },
        { "type": "image", "image": "train/bingo_2100/Bingo_2100_bingo1/screenshots/screenshot_step_0.png" }
    ]},
    { "role": "assistant",
      "content": "<think>I read the 3x3 grid cell by cell: (0,0) rhino, (0,1) cheetah … </think> …action…" }
    // multi-step trajectories continue with further user(screenshot)/assistant(<think>+action) turns
  ],
  "tools": [ /* Computer-Use tool schema: screenshot, click, drag, type_text, hold */ ]
}

tools declares the five Computer-Use tools available to the agent: screenshot, click, drag, type_text, hold (submitting is a click on the page's Submit button — there is no separate submit tool).

Loading

After your access request is approved, log in and download (jsonl + screenshots):

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

Grab a single task (e.g. just Bingo train), or only the .jsonl without images:

from huggingface_hub import snapshot_download
# one task, with its screenshots
snapshot_download("ZHEN-04/CaptchaArena-Trajectories", repo_type="dataset",
                  allow_patterns=["train/bingo_2100/*", "train/bingo_sft_2100_thinking_perturn.jsonl"],
                  local_dir="CaptchaArena-Trajectories")
# all training jsonl only (no images)
snapshot_download("ZHEN-04/CaptchaArena-Trajectories", repo_type="dataset",
                  allow_patterns=["train/*.jsonl"], local_dir="CaptchaArena-Trajectories")

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-Trajectories, please cite:

@article{zhang2026captchaarena,
  title   = {CaptchaArena: A Large-Scale, Fine-Grained Dataset for Training Computer-Use Agents on Interactive CAPTCHAs},
  author  = {Zhang, Zhenhao and Fan, Zhaoyu and Ying, Haohan and Hu, Jingwen and Fan, Hancen and Zhou, Junhao and Chen, Zitian and Zhu, Linchao},
  journal = {arXiv preprint arXiv:2609.31957},
  year    = {2026}
}

Credits

The web environment these trajectories were recorded in began as Open CaptchaWorld (Luo et al., NeurIPS 2025): its page template, layout, frontend code and parts of its content are reused here under its MIT license. This is why every screenshot shows the "Open CaptchaWorld" page title rather than a CaptchaArena name. The trajectories, their replay-verified actions and the reasoning annotations in this repo were produced for this project.

If you use this dataset, please also cite Open CaptchaWorld:

@inproceedings{luo2025opencaptchaworld,
  title     = {Open CaptchaWorld: A Comprehensive Web-based Platform for Testing and Benchmarking Multimodal LLM Agents},
  author    = {Luo, Yaxin and Li, Zhaoyi and Liu, Jiacheng and Cui, Jiacheng and Zhao, Xiaohan and Shen, Zhiqiang},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {38},
  year      = {2025}
}

Downloads last month: 94

Total file size:

51.9 GB

Models trained or fine-tuned on ZHEN-04/CaptchaArena-Trajectories

ZHEN-04/CaptchaAgent

Image-Text-to-Text • Updated 11 days ago • 2

Paper for ZHEN-04/CaptchaArena-Trajectories

CaptchaArena: A Large-Scale, Fine-Grained Dataset for Training Computer-Use Agents on Interactive CAPTCHAs

Paper • 2609.31957 • Published 16 days ago • 6