About Continual Learning

[evi-Genius]

hi, it's a impressive work! But I have some questions. When training fluency, what's the script would be like? Should the _model_name_orpath and _outputdir will be changed into _continual_summcoherence and _continual_summfluency ? Also, how many samples per phase? 30000 for the coherence ,36000 for fluency ,42000 for consistency and 48000 for relevance? It's would be much helpful if you could release some script and code about it!

[maszhongming]

Thanks for your interest in our work! Yes, your understanding is right. You'll need to update both model paths, and the number of data you provided is correct.

Please note that the training steps needed may vary on the learning difficulty of the dimension. For example, achieving the best performance for fluency and coherence typically takes about 1 epoch, while for consistency and relevance, it may take less than 1 epoch.

[evi-Genius]

Thanks for your interest in our work! Yes, your understanding is right. You'll need to update both model paths, and the number of data you provided is correct.

Please note that the training steps needed may vary on the learning difficulty of the dimension. For example, achieving the best performance for fluency and coherence typically takes about 1 epoch, while for consistency and relevance, it may take less than 1 epoch.

Thanks for your reply! It's very useful!