Here are instructions for training a model using the Colab notebook script.
At the Clone repo section, put in GitHub user and access token in indicated locations.
In the first few lines after Train transformer, there are hyperparameters that can be changed for training, including epochs and batch size.
When the first cell is run, follow the prompt for authorization.
The Colab script does the following:
- Copy dataset from gcp bucket to Colab environment
- Clone repo
- Train model
- Save checkpoints on gcp bucket while training
- Evaluate every checkpoint after training
- Update desired google sheet with evaluation results