PixelProse Dataset Overview

Basic Information

• License: cc-by-4.0
• Task Categories:
  • Image‑to‑Text
  • Text‑to‑Image
  • Visual Question Answering
• Language: English
• Tag: croissant
• Name: PixelProse
• Size Category: 10M<n<100M

Configuration

• Default Config:
  • Training Set: data/vlm_captions_*.parquet
  • CC12M: data/vlm_captions_cc12m_*.parquet
  • CommonPool: data/vlm_captions_common-pool_*.parquet
  • RedCaps: data/vlm_captions_redcaps_*.parquet

Details

• Total Image‑Text Pairs: 16,896,214 (16.9M)
  • CommonPool: 6,538,898 (6.5M)
  • CC12M: 9,066,455 (9.1M)
  • RedCaps: 1,290,861 (1.3M)

Data Download

• Parquet Files:
  • Via Git LFS:

    git lfs install
    git clone https://huggingface.co/datasets/tomg-group-umd/pixelprose
    

  • Via HuggingFace API:

    from datasets import load_dataset
    ds = load_dataset("tomg-group-umd/pixelprose")
    

  • Direct Link: access the data directory to download required files.

Columns

• uid: unique image identifier
• url: image URL
• key: image‑related key
• status: status returned by vlm_model
• original_caption: original inherited caption
• vlm_model: model used for captioning
• vlm_caption: dense caption from PixelProse
• toxicity: score for general harmful behavior
• severe_toxicity: score for extremely harmful or abusive language
• obscene: score for obscene or inappropriate language
• identity_attack: score for language targeting individuals or groups based on identity
• insult: score for language meant to insult or demean
• threat: score for language conveying threats of harm
• sexual_explicit: score for language containing explicit sexual content
• watermark_class_id: watermark class (0 = watermarked image, 1 = non‑watermarked, 2 = non‑watermarked with text)
• watermark_class_score: prediction scores for each watermark class, range [0, 1]
• aesthetic_score: aesthetic rating, range [0, 10]
• error_message: error message returned by vlm_model
• width / height: image dimensions used for running vlm_model
• original_width / original_height: original image dimensions
• exif: EXIF metadata of the image file
• sha256: SHA256 hash of the image file
• image_id, author, subreddit, score: attributes inherited from RedCaps (not available in CC12M and CommonPool)