WebDataset export for PACO-style localized caption data.
Samples: 1500
Shards: 2
Payload format inside each shard: pickle
Split: train
Config: PACO
Media files are stored in WebDataset tar shards.
Each sample key is stable and becomes key in the dataset viewer.
Hugging Face will infer columns such as jpg, pickle, json, key, and url from the shard contents.
Manifest file: PACO/annotations.json
