license: apache-2.0 | |
tags: | |
- llava | |
datasets: | |
- Ejafa/ye-pop | |
pipeline_tag: image-text-to-text | |
A ViT-B/32 CLIP model trained for 4 epochs on the [ye-pop](https://huggingface.co/datasets/Ejafa/ye-pop) dataset (491,520 images and [LLaVA 1.5](https://github.com/haotian-liu/LLaVA)-generated detailed captions). Research artifact of [clip-synthetic-captions](https://github.com/nopperl/clip-synthetic-captions). Outperforms the CLIP model trained using the original alt-texts on the [DataComp benchmark suite](https://datacomp.ai) (38 image classification and retrieval tasks). | |
Note: likely not directly useful as it is severely undertrained. | |