philschmid/habana-xlm-r-large-amazon-massive

This model is a fine-tuned version of xlm-roberta-large on the AmazonScience/massive dataset. It achieves the following results on the evaluation set:

8x HPU approx. 41min

train results

{'loss': 0.2651, 'learning_rate': 2.4e-05, 'epoch': 1.0}
{'loss': 0.1079, 'learning_rate': 1.8e-05, 'epoch': 2.0}
{'loss': 0.0563, 'learning_rate': 1.2e-05, 'epoch': 3.0}
{'loss': 0.0308, 'learning_rate': 6e-06, 'epoch': 4.0}
{'loss': 0.0165, 'learning_rate': 0.0, 'epoch': 5.0}

total

{'train_runtime': 3172.4502, 'train_samples_per_second': 127.028, 'train_steps_per_second': 1.986, 'train_loss': 0.09531746031746031, 'epoch': 5.0}

eval results

{'eval_loss': 0.3128528892993927, 'eval_accuracy': 0.9125852013210597, 'eval_f1': 0.9125852013210597, 'eval_runtime': 45.1795, 'eval_samples_per_second': 314.988, 'eval_steps_per_second': 4.936, 'epoch': 1.0}
{'eval_loss': 0.36222779750823975, 'eval_accuracy': 0.9134987000210807, 'eval_f1': 0.9134987000210807, 'eval_runtime': 29.8241, 'eval_samples_per_second': 477.165, 'eval_steps_per_second': 7.477, 'epoch': 2.0}
{'eval_loss': 0.3943144679069519, 'eval_accuracy': 0.9140608530672476, 'eval_f1': 0.9140
608530672476, 'eval_runtime': 30.1085, 'eval_samples_per_second': 472.657, 'eval_steps_per_second': 7.407, 'epoch': 3.0}
{'eval_loss': 0.40938863158226013, 'eval_accuracy': 0.9158878504672897, 'eval_f1': 0.9158878504672897, 'eval_runtime': 30.4546, 'eval_samples_per_second': 467.286, 'eval_steps_per_second': 7.322, 'epoch': 4.0}
{'eval_loss': 0.4137658476829529, 'eval_accuracy': 0.9172932330827067, 'eval_f1': 0.9172932330827067, 'eval_runtime': 30.3464, 'eval_samples_per_second': 468.952, 'eval_steps_per_second': 7.348, 'epoch': 5.0}

Environment

The training was run on a DL1 instance on AWS using Habana Gaudi1 and optimum.

see for more information: https://github.com/philschmid/deep-learning-habana-huggingface