[QUESTION] finetuning COMET base models

[aenaliph]

Hi,

I am trying to do some ablation studies on my custom pSQM dataset. I am constrained for compute and memory and cannot train the large (L/XL) or explainable models yet.

I would like to do the following:

1. train a DA model from scratch using my own data: I use the config as defined in the regression_model.yaml config file.
2. finetune wmt22-comet-da using the regression_model.yaml config file
3. train a QE model from scratch using my own data: I use the config as defined in the referenceless_model.yaml config file.
4. finetune wmt22-cometkiwi-da using the unified_metric.yaml config file with `input_segments:
   • mt
   • src`

My question is whether 3 and 4 are analogous to 1 and 2? Or should I be training and finetuning the QE models as per the unified_metric.yaml config?

Would appreciate any pointers regarding this.