5fa1a76
1
For tasks such as classification, this is not a problem, but for tasks like masked language modeling or token classification, we need a hidden state with the same sequence length as the original input.
