Data Selection Through Iterative Self-Filtering for Vision-Language Settings

The article presents a novel iterative self-filtering method for data selection in training vision-language models, leveraging a CLIP model. This approach dynamically refines the training dataset by balancing high-probability clean samples with diverse examples, resulting in improved downstream performance without requiring additional or pre-trained data. This method is significant for practitioners as it enhances model training efficiency and effectiveness in handling noisy datasets.