Writing logs to /p/qdata/jm8wx/research/text_attacks/textattack/outputs/training/roberta-base-glue:mrpc-2020-06-29-14:01/log.txt. Loading nlp dataset glue, subset mrpc, split train. Loading nlp dataset glue, subset mrpc, split validation. Loaded dataset. Found: 2 labels: ([0, 1]) Loading transformers AutoModelForSequenceClassification: roberta-base Tokenizing training data. (len: 3668) Tokenizing eval data (len: 408) Loaded data and tokenized in 15.40734577178955s Training model across 4 GPUs ***** Running training ***** Num examples = 3668 Batch size = 16 Max sequence length = 256 Num steps = 1145 Num epochs = 5 Learning rate = 3e-05 Eval accuracy: 88.23529411764706% Best acc found. Saved model to /p/qdata/jm8wx/research/text_attacks/textattack/outputs/training/roberta-base-glue:mrpc-2020-06-29-14:01/. Eval accuracy: 89.2156862745098% Best acc found. Saved model to /p/qdata/jm8wx/research/text_attacks/textattack/outputs/training/roberta-base-glue:mrpc-2020-06-29-14:01/. Eval accuracy: 91.17647058823529% Best acc found. Saved model to /p/qdata/jm8wx/research/text_attacks/textattack/outputs/training/roberta-base-glue:mrpc-2020-06-29-14:01/. Eval accuracy: 88.48039215686273% Eval accuracy: 89.4607843137255% Saved tokenizer to /p/qdata/jm8wx/research/text_attacks/textattack/outputs/training/roberta-base-glue:mrpc-2020-06-29-14:01/. Wrote README to /p/qdata/jm8wx/research/text_attacks/textattack/outputs/training/roberta-base-glue:mrpc-2020-06-29-14:01/README.md. Wrote training args to /p/qdata/jm8wx/research/text_attacks/textattack/outputs/training/roberta-base-glue:mrpc-2020-06-29-14:01/train_args.json.