Most Warzone "datasets" I see are dumps. Random screenshots, no zone state, no timing, no consistency between frames. You can train a classifier on that. You cannot trust the output.

Safer-zone prediction is a temporal problem. The model needs:

• The tac-map as the player saw it

• Where the current zone actually is

• Where it is about to go

• A stable label schema across matches, not vibes

If those four aren't locked, you're fitting noise. The network will look confident. It will still be wrong on the next circle.

That's why CircleSight is labeled frames, not a highlight reel. If you're building game-AI or CV on this, start with the labels, then worry about architecture.

Questions on schema or split strategy — drop them here.