What should I do if I want a blank, untrained ColBRET?

[jmnian]

I have some data. My goal is to train from a vanilla version of ColBERT that uses the bert-base-uncased as the encoder, and the linear layer parameters are randomly initialized. I would love to leverage RAGatouille's utilities for the ease of training. Thank you for pointing me to the right direction.
When initializing the trainer I'm using this:
trainer = RAGTrainer(model_name=model_name, pretrained_model_name = "colbert-ir/colbertv1.9", language_code='en', n_usable_gpus=1 )

Does colbertv1.9 mean I am using an untrained ColBERT? Is there any way I can use RAGatouille to accomplish my goal?

[bclavie]

Hey! The way to start the trainer from a completely untrained ColBERT would be just specifying whatever base model you want to use as an argument to pretrained_model_name. So in your case, you'd run it as pretrained_model_name='bert-base-uncased'.