The Three Phases Of R1-Zero-To-R1

DeepSeek-R1's full pipeline has three headline phases after R1-Zero: _____(1) SFT on curated CoT, _____(2) with GRPO, and _____(3) with a general reward. Fill the 3 blanks. Flag format: CONGRESS{1:[word-with-hyphen],2:[word-with-hyphen],3:[word-with-hyphen]}. Example: CONGRESS{1:cold-start,2:reasoning-rl,3:rejection-sampling}.