I am trying to modify Huggingface Seq2SeqTrainer() and other APIs such that the model works with prefix allowed function. I wrote a prefix_allowed_funtion() to allow different tasks for each examples during train & test phases. However, since start token is always , the model doesn't really make use of prefix_allowed_function() that I wrote. So I want to change the the start_token_id conditioned on the example task, so that the generate() step always know what to generate even from index 0. Please let me know if there's any such techniques available so.
Asked
Active
Viewed 6 times
