Cannot Replicate xlm-roberta-large-xnli Results

[SimoneFaricelli]

Hi Joe! First of all, i'd like to thank you regarding your hard work about BERT and relative models.

But here's the issue: I'm trying to replicate results of the xlm-roberta-large-xnli model, obtained within the HuggingFace Model Page, but it seems impossible: values i get through my python code are always different from the above-mentioned ones. I also tried your notebook, but they're different too.

It's not worth mentioning the specific Sentence and Labels, since ANY string and labels, in any language, i tried, got different results from that page.

Here's the code i used:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
from transformers import pipeline
import torch

device = 0 if torch.cuda.is_available() else -1
tokenizer = AutoTokenizer.from_pretrained("joeddav/xlm-roberta-large-xnli")

model = AutoModelForSequenceClassification.from_pretrained("joeddav/xlm-roberta-large-xnli")

sequence_to_classify = "Con la mia macchina del caffe e la capsula prima esce solo acqua, poi si sente che la capsula viene bucata e infine esce il caffè. Ma intanto l' acqua è nella tazzina. Voglio un rimborso per tutti i soldi che ho speso."
candidate_labels = ["tecnologia", "cibo", "bevande", "finanza", "cinema", "giochi"]

classifier = pipeline("zero-shot-classification",
                      model=model, tokenizer=tokenizer, device=device)

#hypothesis_template = "This text is about {}."
classifier(sequence_to_classify, candidate_labels, hypothesis_template, multi_class=False)

I even tried to use hypotesis_template which increased the values difference even more.