CC3M: Conceptual Captions with 3.3M Web-Harvested Image-Text Pairs

Approximately 3.3 million images annotated with captions harvested from web image alt-text attributes. The dataset was created to provide a wider variety of caption styles compared to curated datasets. It is hosted on Hugging Face by author 'chaocq' and was last updated on March 17,我们发现了一个错误。输入中的最后更新日期是2026-03-17，这明显是一个未来的日期，可能是数据录入错误。根据事实性协议，对于这种明显错误，我们应直接陈述输入中的事实，不做推断或修正。因此，在摘要中应直接使用该日期。

Use Cases

• Train image captioning models based on the large-scale, web-harvested image-text pairs.
• Benchmark vision-language understanding models based on the diverse, non-curated caption styles.
• Study the characteristics and noise patterns of web-scraped alt-text data.
• Pre-train multimodal foundation models on a large corpus of image and text data.

Strengths

• Contains approximately 3.3 million data points, providing substantial scale.
• Captions are harvested from web alt-text, likely representing a wider variety of styles than curated datasets.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown beyond the approximate total, which may limit suitability assessment.
• Data may reflect source bias inherent to web-harvested content from unspecified regions.

Provenance

Source

Web images and their associated alt-text HTML attributes.

Collection Method

Harvested from the web.

Time Range

null

Freshness

Last updated 2026-03-17 05:50:30; freshness should be verified due to the future date.

Geography

null