A new word recognition dataset similar to MNIST, containing 30 words extracted from the Esposalles Dataset. The word images are not binarised, instead the background has been removed as much as possible leaving the strokes unchanged. The dataset is deliberately very imbalanced and some groups of words are very similar, making this dataset very challenging! It has not been split in beforehand making it possible to try different datasplits. Matlab code for reading the words file and the labels file are provided. These are based on code found on the internet that I was not able to find the origin for.