Paper summary
Paper summary: DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
DeepSeekMath 7B continues pre-training on 120B maths tokens mined from Common Crawl and introduces GRPO, a PPO variant without a value network, reaching 51.7% on the MATH benchmark.
- GRPO replaces PPO's learned value baseline with the average reward of a group of sampled answers, which saves memory.
- A carefully filtered web corpus turned out to be a rich source of maths data; arXiv helped less than expected.