YFCC15M Image and Page URL Subset for Vision-Language Training

A subset of approximately 15 million image-text pairs from the YFCC100M dataset, curated for training vision-language models. It was prepared by author vishaal27 and uploaded to Hugging Face in January 2024. The dataset provides page URLs and direct image download URLs for each entry.

Use Cases

• Training contrastive image-text models like CLIP using the provided image-download-urls and associated page metadata.
• Benchmarking dataset robustness and quality by analyzing the ~15M subset against the full YFCC100M collection.
• Preprocessing image-text data for model training using the img2dataset tool with the provided csv file of page-urls and image-download-urls.

Strengths

• Approximately 15 million data points, providing a substantial scale for model training.
• Includes both page-urls and image-download-urls, facilitating efficient data downloading and linking.

Limitations

• The exact number of rows, column definitions, and sample data are not provided in the input.
• Image quality, text relevance, and potential biases from the source YFCC100M dataset are not detailed.

Provenance

Source

Subset of the YFCC100M dataset.

Collection Method

Curated for the paper 'Quality Not Quantity: On the Interaction between Dataset Design and Robustness of CLIP'.

Freshness

Last updated on the platform in January 2024.