Trying to find the Marian version used to train Marian MT transformers. This would help me understand the benchmarks for translation times. I see that multiple papers mention various benchmarks for CPU translation :
Marian v1.9 suggests considerably faster translation times than Marian v1.7 and Marian v1.5.
Finding the version used for the huggingface transformers both in the huggingface and Helsinki NLP documentation hasn't been fruitful. Would be really helpful to have this answered.
Those models were trained by @jorgtied as part of the OPUS project. He might know the answer.
@sshleifer Do you know how can I fine tune the models with my specific data?
I just started a PR to support that, but it's still a week away. At the moment, your best bet is to modify summarization/finetune.py.
Trying to find the Marian version used to train Marian MT transformers. This would help me understand the benchmarks for translation times. I see that multiple papers mention various benchmarks for CPU translation :
* https://www.aclweb.org/anthology/D19-5632.pdf - has the benchmarks for the latest update Marian v1.9 * https://www.aclweb.org/anthology/P18-4020.pdf - does not mention the version but I presume it uses versions between Marian v1.7 and Marian v1.5Marian v1.9 suggests considerably faster translation times than Marian v1.7 and Marian v1.5.
Finding the version used for the huggingface transformers both in the huggingface and Helsinki NLP documentation hasn't been fruitful. Would be really helpful to have this answered.
The version should be marked in the original model files distributed at https://github.com/Helsinki-NLP/Opus-MT. Most of them will be v1.7. I just started recently to use v1.9 for upcoming models.
This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.
Most helpful comment
The version should be marked in the original model files distributed at https://github.com/Helsinki-NLP/Opus-MT. Most of them will be v1.7. I just started recently to use v1.9 for upcoming models.