we have detected mislabeling errors in the training set. It affects 70 files all of 1203 Hindi handwritten training files. Around 6% of Hindi handwritten files. You have the list of mislabeled files here: https://dauam-my.sharepoint.com/:t:/g/personal/aythami_morales_uam_es/EaM3vLKYPpFDoU7Xh1GYbt4BbLK-p6oY4RR8_RTHz_7_4A?e=c6Pix6
We recommend removing these images from your training frameworks.