Using the style lora dataset originally trained on SDXL, completely different results were achieved. Not sure why, but it worked to some extent and could be used to test lora usage in sd3.

Below is the effect of this dataset in sdxl