Philipp Schmid • 1/17/2025

Bite: How Deepseek R1 was trained

This technical article details how DeepSeek AI trained its DeepSeek-R1 model, an open model rivaling OpenAI's o1 in reasoning. It explains the Group Relative Policy Optimization (GRPO) algorithm, a reinforcement learning method that eliminates the need for a value function and uses group-based scoring. The summary covers the multi-stage training process, rule-based rewards, and the resulting performance gains in mathematical and coding tasks.

0 comments

#Reinforcement Learning #Deepseek #LLM Training