GRPO

An in-depth technical analysis of Andre Karpathy's 300-line GRPO implementation in nanochat, exploring how simplified reinforcement learning can solve complex reasoning tasks like GSM8K.

Meta FAIR's new Autodata framework proves that data quality, defined by model behavior rather than static rubrics, allows a 4B parameter model to surpass models nearly 100x its size.

DeepSeek recently updated its R1 technical report from 22 to 86 pages, revealing the intricate details of its multi-stage training pipeline, failed experiments, and the path to DeepSeek-V4.