tune_layers

Tuning best layer of a pre-trained English model on WMT16 dataset

Downloading the dataset

This downloads the WMT16 dataset and extracts it into a new folder called wmt16. If the folder wmt16 exists, it will skip the process.

bash download_data.sh

Tuning the models

Here is an example of tuning three models in a row:

python tune_layers.py -m bert-base-uncased roberta-base albert-base-v2

The results would be appended to best_layers_log.txt. The last three lines of best_layers_log.txt would be

'bert-base-uncased': 9, # 0.692518813886652
'roberta-base': 10, # 0.7062886932674598
'albert-base-v2': 9, # 0.6682362357086912

which shows the model name, the best number of layers, and the pearson correlation with human judgement. These can be copied and pasted into model2layers in bert_score/utils.py.