Update README.md

README.md

@@ -17,10 +17,11 @@ Wandb at https://wandb.ai/huseinzol05/nanot5-small-malaysian-cased-translation-v
 
 ## how we trained it?
 
-We done 2 phases,
+We done 3 phases,
 
-1. First phase, trained on 6B tokens noisy translation dataset.
-2. Second phase, trained on 1B tokens higher quality translation dataset.
+1. First phase, trained on 5% of the 6B tokens noisy translation dataset that include all prefixes on padding based training to improve attention bias.
+1. Second phase, trained on 6B tokens noisy translation dataset on packing based and this required to freeze attention bias to speed up the training.
+2. Third phase, trained on 1B tokens higher quality translation dataset on packing based and this required to freeze attention bias to speed up the training.
 
 ## Supported prefix