license: mit
datasets:
- wikipedia
- bookcorpus
language:
- en
metrics:
- glue
library_name: transformers
This is our reproduction using the official HuggingFace roberta architecture with a medium size. On the architecture side, RoBERTa is exactly the same as BERT except for its larger vocabulary size.
According to Google's BERT releases and BERT-Medium, a medium sized model should have a config of Layer=8, Hidden=512, #AttnHeads=8, and IntermediateSize=2048. We follow this config to pre-train a RoBERTa-base model for reproduction.
We use the same datasets as BERT (English Wikipedia and Book Corpus) to pre-train for 30k steps with a batch size of 8,192. I also released the reproduction of this dataset on HuggingFace.
We utilized DeepSpeed ZeRO-2 for performance optimization.
Other training configuration:
| Parameter | Value |
|---|---|
| WARMUP_STEPS | 1800 |
| LR_DECAY | linear |
| ADAM_EPS | 1e-6 |
| ADAM_BETA1 | 0.9 |
| ADAM_BETA2 | 0.98 |
| ADAM_WEIGHT_DECAY | 0.01 |
| PEAK_LR | 1e-3 |
We achieve very similar performance as the official BERT-Medium release on GLUE:
| Model | MRPC-F1 | STS-B-Pearson | SST-2-Acc | QQP-F1 | MNLI-m | MNLI-mm | QNLI-Acc | WNLI-Acc | RTE-Acc |
|---|---|---|---|---|---|---|---|---|---|
| RoBERTa-medium (ours) | 83.6 | 82.7 | 89.7 | 89.0 | 79.7 | 80.1 | 89.3 | 31.0 | 57.4 |
| BERT-medium | 86.3 | 87.7 | 88.9 | 89.4 | 80.6 | 81.0 | 89.2 | 29.6 | 63.9 |
Evaluation Scores Curve (AVG of scores) during pretraining:
For both stats above we don't report CoLA scores as it's pretty unstable. The raw CoLA scores are:
| Step | 1500 | 3000 | 6000 | 9000 | 13500 | 18000 | 24000 | 30000 |
|---|---|---|---|---|---|---|---|---|
| CoLA | 1.7 | 13.5 | 29.2 | 31.4 | 31.1 | 24.1 | 29.0 | 20.0 |