Creating summaries of fixed length with PEGASUS model

Hi all,

I am using a Pegasus model (google/pegasus-multi_news) to generate summaries from new articles and trying to get short summaries (20-50 words).

model_name = 'google/pegasus-multi_news'
model = PegasusForConditionalGeneration.from_pretrained(model_name)
model.generate(**batch, **kwargs)

I tried adjusting the max_length, length_penalty, early_stopping arguments but they don’t change the output for me.

max_new_tokens does give me a shorter output but it is essentially truncated text so not that useful.

Is there a way to force/strongly guide the model to generate a summary of roughly the desired length?

I’m afraid there’s no way to do this in Pegasus, or any abstractive summarisation model.