CLIP-ViT-B-32-DataComp.XL-s13B-b90K
| Property | Value |
|---|---|
| License | MIT |
| Framework | OpenCLIP |
| Paper | DataComp Paper |
| Training Data | DataComp-1B (1.4B samples) |
What is CLIP-ViT-B-32-DataComp.XL-s13B-b90K?
This is a Vision Transformer model trained using the CLIP framework on the massive DataComp-1B dataset. The model utilizes a ViT-B/32 architecture and was trained at stability.ai. It achieves an impressive 72.7% zero-shot top-1 accuracy on ImageNet-1k, making it particularly powerful for zero-shot image classification tasks.
Implementation Details
The model is built on the OpenCLIP framework and trained with 1.4 billion samples from the DataComp-1B dataset. It's designed for efficient processing of image-text pairs and demonstrates strong performance across 38 different evaluation datasets.
- Architecture: Vision Transformer Base model with 32x32 patches
- Training Infrastructure: stability.ai cluster
- Dataset: DataComp-1B with comprehensive evaluation suite
Core Capabilities
- Zero-shot image classification
- Image and text retrieval
- Foundation for downstream task fine-tuning
- Image generation guidance and conditioning
Frequently Asked Questions
Q: What makes this model unique?
This model stands out due to its training on the carefully curated DataComp-1B dataset and its impressive zero-shot classification capabilities, achieving 72.7% accuracy on ImageNet-1k without any fine-tuning.
Q: What are the recommended use cases?
The model is primarily intended for research purposes, including zero-shot classification, image-text retrieval, and as a foundation for downstream tasks. However, it's important to note that deployed commercial use cases are currently out of scope.
