Hey gang getting errors with PEGASUS PARAPHRASE

Beginners
1

I’m getting these random output errors maybe 1/50 times I process a sentence.

generated text: Is a 12 person tent 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110 888-739-5110

This is the number for some random pizza place. Why does it keep appearing in my output. Is there anyway I can make sure this doesn’t happen again. I can’t find the pattern to my input data causing this.

Please let me know if you have a solution. I have found this error on both the hugging face interference and also using torch on Gcollab. It’s driving me nuts.

2

Hey, I’m also getting the same issues. Were you by any chance able to figure out the bug?

3

I got the same error (also a repetition of the number 888-739-5110) for the following input text: HuggingFace Pegasus XSUM error

After investigating the issue, it seems that the probability vector contained an infinity value. To overcome the issue, I used the following parameters for the model:

parameters = {
            "temperature": 0.1,
            "repetition_penalty": 100.0,
            "do_sample": False,
            "top_k": 100,
            "top_p": 0.8,
            "no_repeat_ngram_size": 3,
}

This will overcome the issue with the repetitive numbers at least (and solved my problem).