Closed yongzx closed 2 years ago
@vnikouliNLE Can you help review XNLI evaluation? I am running into the issue where evaluating right after the training (i.e., have do_train
and do_predict
in a single script at the same time) gives wildly different results than evaluating the check-points only (first run do_train
then run do_predict
.)
I have made the following changes:
load_model
function intoload_task_specific_adapters
,load_embedding_layers
, andload_language_adapters
.do_eval_after_train
argument. Simply usedo_predict
will do.baseline
argument for simply fine-tuning the BLOOM model without language adapters and replacing embedding layers. May be redundant and to be removed in the future.