Mistral 7b v0.2 with attention_dropout=0.6, for training purposes