VidaForge: Open Research Infrastructure for Video Pretraining Data Recipes
VidaForge makes video pretraining data recipes executable and inspectable, then releases a 3.14 million-clip dataset for studying them.
The system models a recipe as a five-stage workflow from raw video to training data, with provenance preserved for each sample. The authors test recipe variants on early from-scratch pretraining for Wan 2.1 and V-JEPA 2.1. Broader coverage produced the best downstream benchmark scores across both objectives, while loss-based evaluation preferred different recipes. HF Daily Papers' note
The system models a recipe as a five-stage workflow from raw video to training data, with provenance preserved for each sample. The authors test recipe variants on early from-scratch pretraining for Wan 2.1 and V-JEPA 2.1. Broader coverage produced the best downstream benchmark scores across both objectives, while loss-based evaluation preferred different recipes. HF Daily Papers' note
score 5