BartForCausalLM vs BartForConditionalGeneration

My questions is when should you use one or the other.

BartForConditionalGeneration states it is good for summarization, however what about other generative tasks (for example modeling a conversation)? On one hand a response should be conditioned on the previous text, but on the other it’s should include additional text that is not just a summarization of the previous text.

2 Likes