Datasets

torch_world.datasets holds the data the models train on: windows over collected episodes for world models, and medical images and questions for vision-language models. Every dataset is a standard torch.utils.data.Dataset that returns dicts of tensors, so it plugs into a DataLoader as is.

• Trajectories

  ---

  Sliding windows over episodes collected in any environment, for training world models.

  TrajectoryDataset

• MedMNIST

  ---

  Five MedMNIST v2 image classification datasets, downloaded and checked on first use.

  MedMNIST

• Medical VQA

  ---

  Templated visual question answering over MedMNIST, in the style of the MMBU challenge.

  MedicalVQA

At a glance

dataset	each item	source	download	used by
TrajectoryDataset	a window of pixels, actions and states	any environment, or your own episodes	none	LeWM, DINO-WM
MedMNIST	an image and its label	MedMNIST v2, CC BY 4.0	Zenodo, MD5-checked	image classification, linear probes
MedicalVQA	an image, its caption, a question and its answer	MedMNIST v2 + question templates	Zenodo, MD5-checked	VL-JEPA

Downloads are cached in ~/.cache/torch_world/ unless you pass another root.