Unveiling the World's Largest LLM Data Set: 3T Tokens of Open-Source Language Models