Hello,
As part of a Master's Thesis project at the University of Geneva, we're investigating REBEL's performance on a variety of
datasets, and we're struggling to understand one aspect of your (brilliant!) paper.
In the original REBEL article, the conclusion states the following :
"We make REBELpre−training available as a standalone RE for more than 200 relation types together with a pre-trained RE model to serve as a baseline when fine-tuning on new RE datasets. "
However, in the hugging face repository, the only model available seems to be rebel-large, what is this "pre-trained RE model" that can serve as a baseline ? I'm confused about this, in the paper too, you make the distinction between $REBEL$ and $REBEL_{pre-training}$ in Table 2 for example,
As far as I understand it, $REBEL$ here is the same model as rebel-large on hugging face, and is just the REBEL seq2seq model
trained on the REBEL dataset, and $REBEL_{pre-trained}$ refers to the fine-tuned REBEL model on the particular dataset
in whose column we are, e.g. fine tuning REBEL on NYT dataset yields an F1 score of 92.0, and this is in the line $REBEL_{pre-training}$ is my understanding correct ?
Best regards and many thanks for any pointers,
A. Freeman
Hello,
As part of a Master's Thesis project at the University of Geneva, we're investigating REBEL's performance on a variety of
datasets, and we're struggling to understand one aspect of your (brilliant!) paper.
In the original REBEL article, the conclusion states the following :
"We make REBELpre−training available as a standalone RE for more than 200 relation types together with a pre-trained RE model to serve as a baseline when fine-tuning on new RE datasets. "
However, in the hugging face repository, the only model available seems to be$REBEL$ and $REBEL_{pre-training}$ in Table 2 for example,
rebel-large, what is this "pre-trained RE model" that can serve as a baseline ? I'm confused about this, in the paper too, you make the distinction betweenAs far as I understand it,$REBEL$ here is the same model as $REBEL_{pre-trained}$ refers to the fine-tuned REBEL model on the particular dataset$REBEL_{pre-training}$ is my understanding correct ?
rebel-largeon hugging face, and is just the REBEL seq2seq modeltrained on the REBEL dataset, and
in whose column we are, e.g. fine tuning REBEL on NYT dataset yields an F1 score of 92.0, and this is in the line
Best regards and many thanks for any pointers,
A. Freeman