Num_return_sequences Parameter in inference api

Hi guys,

I have the following problem, I am trying to use various paraphrasing models ( I have used the pegasus paraphraser and also some t5 paraphrasers) by using the inference api.

Sadly the parameter num_return_sequences doesn’t seem to work with these models. The response I get always only contains one single paraphrase. I would like to produce 3-4 pharaphrased sentences to my input.

I found that the num_return_sequences parameter does work when using gpt2 or gpt-j, but these models often produce some low quality paraphrases, that’s why I wanted to use the Pegasus paraphraser instead.

Would appreciate it if somebody knows how I can produce multiple paraphrases for one single input using the inference api

Yes. Pegasus models only return one result on the inference api. Changing the num_return_sequences seems to have no effect. Anyone know why this is?

Anyone know why the num_return_sequences doesn’t work on any of the paraphrasing models in the inference api?