CaptchaArena-Trajectories 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.

⏳ Access is not open yet. We will start sharing CaptchaArena-Trajectories 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.

CaptchaArena-Trajectories

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 families.

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 / submit …), observes the next screenshot, and repeats until the puzzle is solved.

⚠️ 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 — puzzle generation, mock-solving, and evaluation.

Data generation pipeline

Trajectories were produced by a generate → verify → repair → human-check loop:

Stage	Model / actor	Role
1. Generation	GPT-5.4-mini	Generates only the CoT `<think>` reasoning for each step. The Computer-Use actions (ground truth) are authored by humans + the benchmark program, not by the model.
2. Verification	Gemini 2.5 Flash × 2	Two independent verifiers check reasoning quality — no information leakage, no perception errors (e.g. mistaking one animal/object for another), and that the `<think>` is consistent with the target action (e.g. if the answer is to click the fox, the reasoning must not describe it as something else).
3. Repair	GPT-5.5 → human	Trajectories that fail verification have their reasoning regenerated by GPT-5.5; if GPT-5.5 still cannot fix it, a human writes it.
4. Human check	Human annotators	Final manual review before a trajectory is admitted to the dataset.

Only trajectories that pass verification (or repair) and human check are included.

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

Families differ in interaction style and trajectory length: some are single-step (e.g. Geometry_Click, Hold_Button, Pick_Area), others are long multi-step rollouts (e.g. Patch_Select, Click_Order, Rotation_Match).

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)	—	—	169,945 (≈ 50.9 GB)

2,100 puzzles per task in Train, 200 per task in Val, across all 20 families.
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. This is why the number of .jsonl lines exceeds the number of puzzles, and why the expansion factor varies by family (single-step families ≈ 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 action schema (click, drag, hold, type, submit, …) */ ]
}

CoT reasoning is carried in the assistant turn inside <think>…</think>, followed by the action.
Multimodal: user turns embed the current screenshot as an image content part referencing a relative path.
tools declares the Computer-Use action space available to the agent.

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")

After downloading, resolve each image path relative to the local repo root.

Relation to CaptchaArena

ZHEN-04/CaptchaArena — the puzzles (images + instruction + ground truth).
ZHEN-04/CaptchaArena-Trajectories (this repo) — CoT Computer-Use agent trajectories that solve those 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-Trajectories, please cite this repository and the code at github.com/X0X0X00/CaptchaArena. (Formal citation / paper reference: TODO.)

Downloads last month: 13

Total file size:

51.9 GB