← Dataset panel
Showing 168 benchmark candidates auto-filtered from the
Action100M-preview
(HowTo100M source). Each video was kept by GPT-4o judging its hierarchical Action100M
captions — see the "LLM Prompt" panel for the exact prompt.
LLM Prompt (sent to GPT)
System prompt
User message (compact video summary)
Raw LLM output JSON
Tree-of-Captions (ground truth from Action100M)