Task generation by VLM oracle

[creativeessence]

Generate synthetic tasks from a randomly chosen video using Oracle VLM instead of using ActivityNet Captions dataset

[IlyasFardaouix]

I'm a bit confused about the task generation process - are we switching from ActivityNet Captions to a random video chosen from the VLM oracle, or are we using the VLM oracle to generate captions for a randomly chosen video? Either way, I'd love to see the code path where the VLM oracle is integrated, is it in a specific function or file that we should be looking at?

[creativeessence]

• Choose a video online randomly
• Randomly select timestamp interval
• Generate description for the chosen timestamp interval
• Compose a synthetic task using this (interval, description)