Tag: DeepSeek
-
GRPO vs PPO for LLM Reasoning: DeepSeek-R1 Explained
How DeepSeek-R1 trains chain-of-thought reasoning with GRPO instead of PPO. Group Relative Policy Optimization removes the critic network, cuts memory by 40%, and produces emergent self-verification behaviors. Full algorithm breakdown with PyTorch code.
-
MoE ์ํคํ ์ฒ: Mixtral๋ถํฐ DeepSeek-MoE๊น์ง ์์ ๋ถ์
MoE architecture explained: how Mixtral and DeepSeek-MoE achieve 8x parameters with 2x compute. Implementation guide included.