README.md

metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: label
      dtype: string
  splits:
    - name: train
      num_bytes: 45122565
      num_examples: 20000
    - name: validation
      num_bytes: 674338
      num_examples: 300
    - name: test
      num_bytes: 1129219
      num_examples: 500
  download_size: 37814659
  dataset_size: 46926122
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*

Dataset Summary

I have uploaded previously a dataset similar to this one, and that's why this one is named with the suffix _v2. In this dataset card, we shall refer to the previous dataset by the name of v1.

This v2 version attempts to fix the following issues:

• There were simply too many images in v1 for any model to properly run through even a single epoch.
  • Consequently, maybe by bad luck, someone may train on images that do not cover all the symbols.
• v1 seems to lack of CAPTCHA images with repeated symbols, e.g. "jj12oj".

The usage and meaning of the current v2 dataset should be intuitive (and quite independent of v1):

In [1]: from datasets import load_dataset

In [2]: dataset = load_dataset("phunc20/nj_biergarten_captcha_v2)
README.md: 100%|█████████████████████████████████████████████████████████████████████████████████████████| 533/533 [00:00<00:00, 1.58MB/s]
train-00000-of-00001.parquet: 100%|██████████████████████████████████████████████████████████████████| 36.3M/36.3M [00:07<00:00, 2.02MB/s]
validation-00000-of-00001.parquet: 100%|███████████████████████████████████████████████████████████████| 541k/541k [00:00<00:00, 2.06MB/s]
test-00000-of-00001.parquet: 100%|█████████████████████████████████████████████████████████████████████| 931k/931k [00:00<00:00, 2.04MB/s]
Generating train split: 100%|████████████████████████████████████████████████████████████| 20000/20000 [00:00<00:00, 113382.55 examples/s]
Generating validation split: 100%|████████████████████████████████████████████████████████████| 300/300 [00:00<00:00, 45083.88 examples/s]
Generating test split: 100%|██████████████████████████████████████████████████████████████████| 500/500 [00:00<00:00, 92186.56 examples/s]

In [3]: dataset
Out[3]:
DatasetDict({
    train: Dataset({
        features: ['image', 'label'],
        num_rows: 20000
    })
    validation: Dataset({
        features: ['image', 'label'],
        num_rows: 300
    })
    test: Dataset({
        features: ['image', 'label'],
        num_rows: 500
    })
})

In [4]: dataset["test"][0]["label"]
Out[4]: '9ymyht'

In [5]: dataset["test"][0]["image"]
Out[5]: <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=140x50>

Citation Information

@ONLINE{nj_biergarten_captcha_v2,
    author = "phunc20",
    title  = "nj_biergarten_captcha_v2",
    url    = "https://huggingface.co/datasets/phunc20/nj_biergarten_captcha_v2"
}