This is a finetuned XLM-RoBERTA model for natural language inference. It has been trained with a massive ammount of data following the ANLI pipeline training. We include data from:

mnli {train, dev and test}
snli {train, dev and test}
xnli {train, dev and test}
fever {train, dev and test}
anli {train}

The model is validated on ANLI training sets, including R1, R2 and R3. The following results can be expected on the testing splits.

Split	Accuracy
R1	0.6610
R2	0.4990
R3	0.4425

Multilabels

label2id={
    "contradiction": 0,
    "entailment": 1,
    "neutral": 2,
},