recaptcha-390k

recaptcha-390k is a dataset of 67,226 reCAPTCHA v2 image challenges containing 390,210 photos in total. Each example includes the original challenge image, the target object (for example, "bus"), and the grid cells that contain that object.

The dataset was used to fine-tune RC1, a set of models that select the matching cells in reCAPTCHA v2 image grids.

Dataset structure

The dataset has two configurations, one for each grid type:

• 3×3: nine separate photos arranged in a grid.
• 4×4: a single photo split into 16 cells.

Each challenge is stored as a single image file.

Configuration	Challenges	Photos	Cells per challenge
3x3	40,373	363,357	9
4x4	26,853	26,853	16
Total	67,226	390,210	—

Targets

The table below shows the number of challenges for each target across all splits, including review and excluded. A dash means the target does not appear in that configuration.

Target	3×3	4×4
Bicycle	5,010	5,468
Bus	10,786	4,731
Car	5,980	—
Chimney	—	13
Crosswalk	6,392	1,144
Fire hydrant	12,181	675
Motorcycle	14	8,998
Parking meter	2	7
Stairs	1	1,809
Taxi	—	108
Tractor	—	161
Traffic light	7	3,739

The targets are heavily imbalanced, and some appear only a few times in a given configuration.

Labels

Cells are numbered starting from 0, left to right and then top to bottom. For a 3×3 grid:

0 1 2
3 4 5
6 7 8

The correct_answers field holds the primary label for each challenge. Labels come from three sources:

Source	Applies to
Claude Opus 5.5	582 challenges (the review split)
Qwen/Qwen3.8-27B	All other 3×3 challenges
facebook/sam3 (SAM 3)	All other 4×4 challenges

Fields

Field	Description
correct_answers	Primary label: indices of cells containing the target
uncertain_cells	Cells in the primary label that are marked as uncertain
label_source	Source of the primary label
machine_answers	Original machine-generated answer, kept even when a review changed it
machine_uncertain_cells	Cells the machine labeler marked as uncertain
machine_label_source	Model that produced the machine-generated answer
reviewed	true if the challenge was reviewed by Claude
exclusion_reason	Why a challenge was placed in the excluded split

Machine-generated labels may contain errors. Some of the Claude-reviewed examples were also used to choose the labeling settings, so keep this in mind when using the review split for evaluation.

Splits

Split	3×3	4×4	Total	Purpose
train	32,047	21,060	53,107	Training
validation	4,006	2,640	6,646	Model selection and hyperparameter tuning
test	3,999	2,657	6,656	Evaluation against machine-generated labels
review	300	282	582	Evaluation against Claude-reviewed labels
excluded	21	214	235	Reference only; not for training or evaluation

The excluded split contains challenges whose machine labels conflicted, as well as challenges that exactly match a review example, either as a whole image or as an individual tile.

In the remaining splits, exact duplicates are always placed in the same split. Near-duplicates are not deduplicated and may appear in different splits.

Usage

Run the following from the directory that contains recaptcha-390k, with Hugging Face Datasets installed:

from datasets import load_dataset
from pathlib import Path
from PIL import Image

data = load_dataset("./recaptcha-390k", "3x3", split="train")
image = Image.open(Path("./recaptcha-390k/data/3x3/train") / data[0]["file_name"])

Replace "3x3" with "4x4" to load the 4×4 configuration.

Each split directory contains the original image files and a metadata.jsonl file. Every metadata record includes a file_name (a path relative to its split directory), the target, the answers, and the label provenance fields. Images are stored as-is, without resizing or recompression.

Hugging Face Datasets loads file_name as a plain string rather than as an image, so you need the image files downloaded alongside the metadata to open them as shown above.

Summary counts are available in dataset_summary.json.

Roadmap

• Add more chimney and parking meter examples.
• Add bridge examples so RC1 can learn to recognize bridges.

Disclaimer

This dataset was created for educational and research purposes. It is not affiliated with or endorsed by Google or reCAPTCHA.

License

The dataset authors' annotations and metadata are released under the MIT License. The source images are not covered by this license. They are included under fair use for research and educational purposes.