Expanded the dataset for training by 50%, and used several training methods, using a preference based merge as the final lora.
Expanded the dataset for training by 50%, and used several training methods, using a preference based merge as the final lora.