COYO-700M

Dataset Overview

Dataset Name

COYO‑700M

Dataset Content

• 747M image‑text pairs, containing multiple metadata attributes to enhance model training.
• Data collected from alt‑text in HTML documents and their associated images.

Data Collection Process

• From October 2020 to August 2021, approx. 10 billion alt‑text‑image pairs were gathered from CommonCrawl.
• Image‑ and text‑level filtering removed non‑informative pairs at minimal cost.

Data Filtering

Image‑Level

• Includes all formats decodable by the Pillow library (JPEG, WEBP, PNG, BMP, etc.).
• Removes images smaller than 5 KB.
• Removes images with aspect ratio > 3.0.
• Removes images with minimum side < 200 px.
• Removes images flagged by OpenNSFW2 or GantMan/NSFW score > 0.5.
• Removes duplicate images based on pHash.

Text‑Level

• Collects English text only.
• Cleans format, discarding texts shorter than 5 characters.
• Discards texts without noun forms.
• Discards texts with word count < 3 or > 256.
• Discards texts appearing more than 10 times.
• Discards texts containing NSFW vocabulary.

Image‑Text‑Level

• Removes duplicate samples based on (image_phash, text).

Dataset Preview

The dataset includes image URLs, text, image dimensions, image pHash, text length, word count, token counts under BERT and GPT models, etc.

Dataset Statistics

• 746,972,269 image‑text pairs.
• 656,114,783 unique URLs.
• 579,679,137 unique image pHashes.
• 566,253,888 unique texts.

Metadata Attributes

• id: Unique 64‑bit integer ID.
• url: Image URL.
• text: Alt text.
• width, height: Image dimensions.
• image_phash: Perceptual hash of the image.
• text_length, word_count: Text length and word count.
• num_tokens_bert, num_tokens_gpt: Token counts for BERT and GPT.
• num_faces: Number of faces detected in the image.
• clip_similarity_vitb32, clip_similarity_vitl14: Image‑text similarity scores using OpenAI CLIP models.
• nsfw_score_opennsfw2, nsfw_score_gantman: NSFW scores for the image.
• watermark_score: Probability of a watermark in the image.