Content Behavior Corpus

The progress of Large Language Models (LLMs) has largely been driven by the availability of large-scale unlabeled text data for unsupervised learning. This work focuses on modeling both content and the corresponding receiver behavior in the same space. Although existing datasets have trillions of content tokens (text, images, audio, and videos), they lack information on receiver effects. To address this, the paper utilizes YouTube, a large publicly available source of content-behavior data, which includes:

Communicator Data: Channel name, and number of subscribers. Message: Youtube video ids, extracted speech, scene-wise captions, on screen text, video description, video length, upload date. Receiver Effect: Video likes, views, and replay graphs.