TL;DR of the paper, generated automatically.

VidUEU-Agent builds instruction data for multimodal models from raw videos, with no hand labeling.

• Pipeline: four stages: filter low-quality shots, mine keyframes, build metadata, synthesize instructions.
• Feedback loop: it retrains on its own results to adjust the data mix. Four rounds, 30K samples each.
• Dataset: VidUEU-100K, 100K samples from 8,268 videos, released on Hugging Face.