Bug: Finetune XLM-RoBERTa-large on XNLI get 0.33 in accuracy while XLM-RoBERTa-base works fine

🤗Transformers
1

Just found something weird: finetuning XLM-RoBERTa-large on XNLI gets 0.33 in accuracy while XLM-RoBERTa-base works fine. I used the “official” script here: transformers/run_xnli.py at main · huggingface/transformers · GitHub