Decreased performance with recent updated model?
Compared with the model when it was first released (the version about 1 month ago), the performance of the new version of the model seems to have dropped a lot, especially in Chinese.
I also found that the Chinese output does not seem to be generated token by token using TextIteratorStreamer, but paragraph by paragraph. Did something go wrong with the sampling?
btw, is there any way to get the previous version of the model?
We didn't alter the model...
I think I might be missing something? Would like to ask is there any difference between "6e6da7b9cdb21eefe6dd8ac9a083554a99d4ce5e" commit and the previous? Because I saw that has updated the .bin file?
No new versions of this particular model have been trained, so I can promise you nothing has changed.
Got it, testing it, thanks for the quick reply
when I update transformers lib to 4.31.0, it can work.
thank!