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

arXiv·medium signal

Nicolicioiu, Ghotra, and Moss propose an iterative self-filtering procedure to select clean, high-value training data at scale for vision-language models, where large clean datasets are otherwise hard to assemble. The method repeatedly filters noisy web-scale data using the model's own signal. Directly actionable for anyone curating training corpora for VLM pretraining or fine-tuning.