Rewriting Pre-Training Data Boosts LLM Performance in Math and Code | Read Paper on Bytez