Closed juliafalcao closed 1 year ago
Yep, Julia you are right. You need first to rescale the data to be between 0 and 1.
The rescaling I used was actually borrowed from BLEURT. The rescaling is the following: 1) Find a reasonable "min value". This is done by finding all annotations with more than 1 annotator where all annotators agreed that the score was 0. Then your "min_value" is the average z-score for those segments. 2) Find a reasonable "max value". This is done by finding all annotations with more than 1 annotator where all annotators agreed that the translation is perfect (100 score). Then your "max_value" is the average z-score for those segments. 3) apply a min_max_scaler to your data and truncate every score above 1 and bellow 0.
This rescaling won't impact your model correlations and your model will output scores in that range.
I am trying to fine-tune the
wmt22-comet-da
model on new DA data, using its checkpoint and the configs available in theconfigs/
folder here. I know that this model scales the result scores between 0 and 1, but after fine-tuning, the new fine-tuned model generates negative scores as well, similarly to the older COMET models.My main question is, where exactly is the configuration I should change for the model to scale the results like
wmt22-comet-da
does? I haven't been able to find that in the code. My configs are exactly the same as those inconfigs/
except for filepaths and the checkpoint to load.