alexixu
alexixu
@AkshitaB The dataset reader is custom class inherit from DatasetReader. Empty instances is ok if the data loader can handle this empty logic. Discarding empty instance is the direct way...
@dirkgr I have tried this, but the default DataLoader implement can not handle None object, And it will throw an Exception of "None type has no index function". I suggest...
@dirkgr I have tried that exactly, by implement _read function return None object. But the Data Loader (not the Dataset Reader) which call the text_to_instance function can not handle None...
I have a similar question that why there is a peak in "loss/total" and "loss/policy"?
@Mryangkaitong No, the KL value becoming bigger and more positive. The policy changes the origin token distribution so the KL value increases.
Thanks! @lvwerra The generation_kwargs has min_length parameter `generation_kwargs = { "min_length": 40} ` This setting will led to wrong KL value, Right? Another question is why short generated text cause...