Zovda
A short video's outcome is largely settled before the caption matters. The opening frames decide whether the rest exists for the viewer, and they are the least recorded part of the dataset.
Most clip datasets describe a video by its topic. Nearly none describe its hook.
What a hook is, as a field
- The first sentence spoken, or the first text shown
- The type of opening: question, claim, demonstration, in-media res
- The time to the first payoff, in seconds
- Whether the opening repeats the title or departs from it
These four predict retention better than the topic tag, and they are cheap to annotate from the frames.
Retention is not engagement
A like is a reaction to the whole. Retention is a curve, and the curve says where people left. A clip with high likes and a cliff at second three is a clip that paid for attention it could not keep.
If you have the curve, store it. If you do not, do not substitute engagement for it.
Format families repeat
Tutorial, reaction, list, story, and demonstration have different rhythm and different failure modes. A dataset that mixes them without a format field learns an average of five things.
Tag the family, count per family, and check the mix before training.
Caption and audio carry different signals
The caption is written for the platform's search. The audio is written for the viewer. They often disagree, and the disagreement is informative.
Zovda assembles these collections through Datuva, where the opening and the format are separate fields. The video and multimodal collection shows the record structure.
Tell us what you need to learn from short video. We reply within one business day.