Hi,

if you haven't resolve the problem, I would suggest reducing the batch_size and the tokenizer max_length (any value below 80 should work for a batch_size of 16 or lower).

Here is something I like to do:

train['length'] = train['Yoruba'].apply(lambda x: len(tokenizer.encode(x)))
train.length.hist()
train.length.describe()

# Do the same for English, and choose your max_length accordingly.