wildworld-dataset — Claude Code skill: work with the WildWorld dataset

What wildworld-dataset does

The skill gives your agent context on WildWorld, a large action-conditioned dataset collected from a photorealistic ARPG. It covers 108M+ frames, 450+ actions, 29 monster species, 4 player characters, 4 weapon types and 5 stages. Every frame carries annotations: character skeletons, states such as HP, animation and stamina, camera poses, depth maps and hierarchical captions.

The skill documents the expected directory layout and Python helpers: a WildWorldClip class for reading a clip and a WildWorldDataset for PyTorch that yields sequences of frames, actions and states. Dependencies are PIL, numpy, torch and torchvision, installed with pip install -r requirements.txt. The dataset and the WildBench benchmark were not yet released at the time of writing, so the code targets a future release. Worth checking the repository before wiring the skill into a pipeline.

Use it when you are preparing data for generative models that predict the next game state from past frames and player actions. If you only need RGB frames, most of the annotation layers will sit unused.