hands-on-modern-rl
An open, still-evolving curriculum that bridges basic reinforcement-learning concepts to LLM alignment, RLVR and agentic RL, with reproducible experiment code and browser-runnable training notebooks.
hands-on-modern-rl is a bilingual, hands-on course covering the path from Markov decision processes and policy optimization through to reasoning models, agents and multimodal systems. It includes reproducible training examples for both agentic reinforcement learning, such as deep-research agents, and traditional reinforcement learning, such as actor-critic continuous control, along with vision-language-model RL experiments.
Online training notebooks, built with ModelScope, let learners run classic reinforcement-learning experiments in a browser without configuring a local environment, and each notebook has a matching companion under the repository's own experiment code. The course was created with AI assistance and is explicitly flagged as not yet fully reviewed, so it may still contain factual or code errors, with a full English translation and PDF builds for both languages produced automatically.