ordergrad
Policy gradient reinforcement learning beyond the mean, e.g., Pass@k, Max@k, TopM@K, CVaR, VaR, Quantiles, Trimmed Means or other functionals of the reward distribution based on unbiased order-statistics gradients. Can be used for LLM training and other REINFORCE / GRPO / PPO based tasks.
Take a look →