Update README.md

README.md

@@ -34,6 +34,11 @@ Note, we are only training 0.58% of the parameters, using only the query, key, a
 trainable params: 23682 || all params: 4075265 || trainable%: 0.5811155838945443
 ```
 
+It was shown in the QLoRA paper that to obtain performance comparable or better than full finetuning, the most important hyperparameter than can 
+that can be adjusted is which weight matrices the LoRA adapters are applied to, with more being better. The rank and other hyperparameters
+such as rank and the scaling factor alpha did not seem to matter. So, an important thing to investigate next would be to check and see if this 
+transfers to protein language models as well. 
+
 ## Testing for Overfitting
 
 ### Checkpoint 1