Congrats to all the winners :D

My approach was quite simple : Single model roberta-base trained on 5 folds for 3 epochs , prediction was the average of all predictions at the last epoch of each fold .

Neither augmentations nor cleaning and preprocessing text helped to improve the score.

Link to the github repo : https://github.com/ahmedattia143/-Basic-Needs-Basic-Rights-Kenya---Tech4MentalHealth-/blob/master/Roerta_Single_Model.ipynb