Datasets:

simulaXrm
/

strokes-opensource

GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/datasets/simulaXrm/strokes-opensource
cd strokes-opensource

git lfs pull \
  --include="deepwriting/parsed/**,deepwriting/split/**,iam/parsed/**,iam/parsed_split/**,iam/split/**" \
  --exclude=""

This restores the same file layout used by the training scripts:

data/
├── deepwriting/
│   ├── parsed/
│   │   └── *.json
│   └── split/
│       ├── train.txt
│       ├── val.txt
│       └── test.txt
└── iam/
    ├── parsed/
    │   └── *.json
    ├── parsed_split/
    │   └── *.json
    └── split/
        ├── train.txt
        ├── val.txt
        └── test.txt

Note:
- The IAM dataset was split into both `iam/parsed/` and `iam/parsed_split/`
  due to Hugging Face directory file-count limits.
- Before training, manually merge the contents of:
  
  iam/parsed_split/ -> iam/parsed/

  so that all IAM JSON files exist under a single `iam/parsed/` directory.
- The `split/` folders contain the dataset split definitions
  (`train.txt`, `val.txt`, and `test.txt`).