For my works, I'm trying to adapt the approch of GENRE in french using BARThez as the initial language model instead of BART.
I think I could make a good Entity Linking model in french (and a great contribution to the community) but my results are currently totally bad.
I think the problem come from BARThez. During my tests, I see it is unable to predict anything (I ask him to give me 5 predictions and it give me only one. The beam parameter isn't respected). I don't think it's because of the tokeniser or the BARThez model but because of the fairseq version
For GENRE, I'm using the fairseq-0.10.0 (it doesn't work with an older version) but you propose the fairseq-0.9.0.
It is highly possible that there is some form of incompatibility which does not prevent the loading or the fine-tuning of BARThez but which makes the result or any attempt at inference unusable.
So, I would like to know :
your opinion with this fairseq problem
if it's possible to share a version of BARThez learned with a newer version of fairseq (0.10.0 or 0.12.0)
if not, if it's possible to share the complete corpus in order to retrain BARThez from scratch with my newer version of fairseq. I could try to retrain it if I have all the data and the parameters for the fairseq-train function.
Hello,
For my works, I'm trying to adapt the approch of GENRE in french using BARThez as the initial language model instead of BART.
I think I could make a good Entity Linking model in french (and a great contribution to the community) but my results are currently totally bad. I think the problem come from BARThez. During my tests, I see it is unable to predict anything (I ask him to give me 5 predictions and it give me only one. The beam parameter isn't respected). I don't think it's because of the tokeniser or the BARThez model but because of the fairseq version
For GENRE, I'm using the fairseq-0.10.0 (it doesn't work with an older version) but you propose the fairseq-0.9.0. It is highly possible that there is some form of incompatibility which does not prevent the loading or the fine-tuning of BARThez but which makes the result or any attempt at inference unusable.
So, I would like to know :
Thanks for your response