Datasets:
Tasks:
Text Generation
Modalities:
Image
Formats:
imagefolder
Languages:
English
Size:
100K - 1M
License:
Search is not available for this dataset
image image |
|---|
End of preview. Expand in Data Studio
Alleviating Attention Bias for Visual-Informed Text Generation
Training Dataset for Allegator finetuning.
The dataset consists of a subset of LLaVA-Instruct-150k and a subset of Flickr30k, with 99,883 and 31,783 samples from each, respectively.
In detail, LLaVA-Instruct-150K contains 158k language-image instruction following samples, including 58k conversations, 23k descriptions, and 77k complex, 182 reasoning. We augment LLaVA-Instruct-150K with Flickr30K, which has relatively short captions, to enrich visual and textual diversity.
Acknowledgement
- Downloads last month
- 6