I1 Captions: Multi-Source Image Captions for Vision-Language Models

I1 Captions is a collection of image-text pairs aggregated from multiple sources, hosted by zlab-princeton on Hugging Face. The dataset comprises several large subsets, including fluxreason (5.9M rows), imagenet22k (13.7M rows), megalith10m (9.4M rows), and others, with the most recent update recorded on May 13, 2026. It provides multiple descriptive captions per image across different subsets.

Use Cases

• Training image captioning models based on the multiple captions (caption1, caption2, etc.) provided per image.
• Benchmarking vision-language alignment using the diverse caption sets from sources like fluxreason and midjourneyv6.
• Fine-tuning large language models for multimodal reasoning tasks based on the structured caption data.
• Conducting comparative analysis of caption quality or style across different source datasets like imagenet22k and inaturalist.

Strengths

• Large scale, with over 13 million rows in the imagenet22k subset alone.
• Provides multiple captions (up to 5) per image in several subsets, offering descriptive variety.
• Aggregates data from several distinct sources, including fluxreason, imagenet22k, and megalith10m, suggesting diversity.

Limitations

• Description metadata is limited; actual data quality, image sources, and caption generation methods require manual inspection after download.
• Column-level documentation is absent; field semantics for columns like 'qwen2vl_2b' or 'no_center_crop' must be inferred.
• Row count for the full dataset is unknown, which may limit suitability assessment for large-scale projects.

Provenance

Source

zlab-princeton

Collection Method

Aggregated from multiple open sources including fluxreason, imagenet22k, and megalith10m.

Freshness

Last updated 2026-05-13 07:00:04; freshness should be verified.