Both tasks are identical, with task 1 using orientation dataset and ask 2 using power dataset. Main distinction between them is for masked lm, task 1 does fine-tuning and evaluatoin on original language whereas task 2 does on English translation of the original text.
Downloads the dataset, chooses the predetermined language and splits the training data. Must be run everytime for the code to be functional.
Uses XLMRoberta, can be run independently of the "Causal LM" section. For the Turkish dataset (biggest), fine-tuning takes approximately an hour with T4.
Uses distilgpt2, can be run independently of the "Masked LM" section. Does a total of 4 evaluations, details can be found in ipnyb files. Fine-tuning is done 2 times, takes up approximately 20 minutes with T4.
Results of the evaluations.