deepseekmath

Definition

A math-specialised open LLM from DeepSeek. Its central contribution is data: a large mathematical corpus filtered out of web crawl and used for continued pre-training, which let a 7B model compete on competition-math benchmarks with much larger general-purpose models. It also introduced GRPO, a reinforcement learning variant that drops the value network and normalises rewards within a group of sampled answers — a recipe later reused for reasoning models.