Is the task string necessary?

[nikogamulin]

Hi,

I have fine-tuned the model to write a compliment for a person, given the person's profile and it works pretty well. In the training examples, I haven't prepended the string 'summarize :' to the source_string column entries. Is it necessary (does it lead to better results) to prepend the string indicating the task?

[Shivanandroy]

For fine-tuning a T5 model, we add “prefix”, to assist multi-task training . In your case, as you are training on single task, so the prefix is not necessary.

Based on my experiments, I found that if you training for a task which T5 was earlier pre-trained on, for e.g. summarization or translation, adding a prefix, might make sense as it is going to converge slightly faster. That's it. But no impact on the results.