Leveraging Image-Text Similarity and Caption Modification for the DataComp Challenge: Filtering Track and BYOD Track
This work addresses the challenge of data design for multimodal AI, offering incremental improvements for researchers and practitioners in the field.
The paper tackled the problem of improving training data quality for multimodal learning by filtering and modifying web crawl data using CLIP and BLIP-2 models, resulting in significant performance gains over DataComp baselines with 6.6% improvement in the filtering track and 48.5% in the BYOD track.
Large web crawl datasets have already played an important role in learning multimodal features with high generalization capabilities. However, there are still very limited studies investigating the details or improvements of data design. Recently, a DataComp challenge has been designed to propose the best training data with the fixed models. This paper presents our solution to both filtering track and BYOD track of the DataComp challenge. Our solution adopts large multimodal models CLIP and BLIP-2 to filter and modify web crawl data, and utilize external datasets along with a bag of tricks to improve the data quality. Experiments show our solution significantly outperforms DataComp baselines (filtering track: 6.6% improvement, BYOD track: 48.5% improvement).