What role does data quality plays in the LLM scaling laws?

Not my post. But still very curious too.