220k GPT4Vision Captions from LIVIS
Multimodal dataset composed of approximately 220,000 images annotated with captions generated by GPT-4Vision from the LVIS dataset. It is intended to improve the understanding and generation of visual annotations.
Approximately 218,000 images and captions, 62.8 MB, Parquet format
Apache 2.0
Description
The dataset 220k GPT4Vision Captions from LIVIS contains images from the LVIS dataset accompanied by captions generated automatically via GPT-4Vision. Captions summarize the original instructions and thus facilitate the training of multimodal image-captioning models.
What is this dataset for?
- Train models for generating image descriptions
- Improving multi-modal contextual understanding of AIs
- Test the ability of models to summarize and formulate accurate captions
Can it be enriched or improved?
It is possible to enrich this dataset by adding human annotations to improve the quality of the captions, or by diversifying the types of summarized instructions for more variety.
🔎 In summary
🧠 Recommended for
- Multimodal AI developers
- Vision & language researchers
- Captioning projects
🔧 Compatible tools
- Hugging Face Transformers
- PyTorch
- TensorFlow
- Multimodal captioning tools
💡 Tip
Complete with human annotations for better quality during fine-tuning.
Frequently Asked Questions
What is the source of the images in this dataset?
The images come from the LVIS dataset, a large dataset of annotated images for object detection.
Are captions generated automatically or manually?
They are generated automatically from LVIS instructions via GPT-4Vision.
Can this dataset be used to train a model commercially?
Yes, the Apache 2.0 license allows commercial use with respect to the conditions.
