Dataset · kdexd
12M image-caption pairs from Reddit
kdexd/red_caps
Large English corpus of 12M images and captions from 350 subreddits for image-to-text tasks.
What's inside
Contains 12 million pairs of images and English text captions. The data was collected from Reddit across 350 manually selected subreddits, whose themes serve as coarse labeling without manual annotation of each example. Images and descriptions cover a wide range of objects and scenes.
What you can do with it
- Pretraining multimodal models and encoders like CLIP
- Training image captioning models
- Fine-tuning vision-language models on diverse everyday scenes
- Tasks
- Image → text
- Languages
- English
- Size
- 10M–100M rows
- Downloads this month
- 3.1K
- Likes
- 60
Models trained on it
- hanxunh/clip_backdoor_rn50_redcaps_clean_labelZero shot image classification 26
- hanxunh/clip_backdoor_rn50_redcaps_blendZero shot image classification 19
- hanxunh/clip_backdoor_rn50_redcaps_nashvilleZero shot image classification 19
- hanxunh/clip_backdoor_rn50_redcaps_badnetsZero shot image classification 15
- hanxunh/clip_backdoor_rn50_redcaps_sigZero shot image classification 14
- hanxunh/clip_backdoor_rn50_redcaps_wanetZero shot image classification 12
How to load it
from datasets import load_dataset
ds = load_dataset("kdexd/red_caps")