Data Filtering for Generative Video Pre-training
Summary
This field-note-style article from Linum describes a data-centric approach to training video models, emphasizing data quality, annotation, and synthetic data. It documents the evolution from CPU-based heuristics to GPU-assisted LLM-based filters, including RLVR, and provides end-to-end pipelines and practical tips for large-scale pre-training. It highlights data filtering as the primary lever to improve model learning.