This repo contains google colab notebooks for finetuning BERT model on Quora Question Pairs dataset. We use colab TPU runtime for faster training. Medium post on this repo can be found here.
If you are not familiar with BERT, please visit The Illustrated BERT, BERT Research Paper and BERT Github Repo.
For experimenting with latest notebook and more details visit colab notebook.
Run in Google Colab
Evaluation results are on BERT base uncased model. We have used train and dev set from GLUE Quora Question Pairs Task.
| Metrics | Train Set | Dev Set |
|---|---|---|
| Loss | 0.150 | 0.497 |
| Accuracy | 0.969 | 0.907 |
| F1 | 0.959 | 0.875 |
| AUC | 0.969 | 0.902 |
| Precision | 0.949 | 0.864 |
| Recall | 0.969 | 0.886 |
Due to limited resources, It was not possible to finetune on BERT Large model and do hyperparamter tuning.