Customizing a trunk model to your needs
You are able to pre-train a model from the Trunk Model tab in the advanced section of the Project Navigation menu or alternatively by clicking on a model within an entity
TEXT MODELS ONLY
At this time Graft supports the training of text models based upon RoBERTa trunk model.
Configuring your new Custom Trunk Model
- Click on CREATE CUSTOMIZED TRUNK MODEL button
If you have clicked through from the Entity Screen the Trunk Model and Data Source will be pre populated and non-modifiable. If you wish to start from a clean configuration please Click on the Trunk Model tab and Click on CREATE CUSTOMIZED TRUNK MODEL
Once on the Configuration Screen update the fields accordingly, Click on FINISH once updated
- Trunk Model which if the Trunk models you wish to pretrain (auto populated from the Entity screen)
- PreTrain Data source which data source contains your data to train the model (auto populated from the Entity screen). The Data source MUST exist ahead of starting this model configuration.
- Display Name the name of your new model
- Data Source Field the field in your data that you wish to use for training the new model
The default Advanced settings are optimized to provide the optimum price/performance balance. Only modify if you understand the impact of the changes.
- Learning Rate A hyperparameter that controls how quickly a model adapts to new data during training
- Number of Epochs to train Number of times the model is trained on the entire dataset, this number has a large effect on training time
- Maximum length of input sequences Truncate the sequence after specified length, this number has a large effect on memory usage