“The third stage of DeepSeek R1's training uses rejection sampling with 800,000 completions, split between 600,000 for reasoning and 200,000 for general chat problems.”