Question about the MLM masking

[longkukuhi]

Hi,

10% of the time, we replace masked input tokens with random word

    indices_random = torch.bernoulli(torch.full(input_ids.shape, 0.5)).bool() & masked_indices & ~indices_replaced

Here is the code you use to replace a token with a random word. Is it correct to use 0.5 as the parameter here?
Thank you for your answer.

[viyjy]

Thanks for your interest in our work.
As shown in this Line, there is ~indices_replaced to make sure only 10% are replaced by random word.
You can reorder this line as:
indices_random = torch.bernoulli(torch.full(input_ids.shape, 0.5)).bool() & ~indices_replaced & masked_indices
Please let me know if you might have any other questions.