Self-hosted alternatives to grpo-rlvr-training

G

grpo-rlvr-training

Train reasoning and verifiable-task behavior

grpo-rlvr-training is an agent skill from the wshobson/agents collection for training reasoning and verifiable-task behavior in language models with GRPO (Group Relative Policy Optimization) and reinforcement learning from verifiable rewards (RLVR). It applies when task success can be checked algorithmically, such as a math answer that is right or wrong, or code that passes its tests, so the model can learn from a reward signal instead of from demonstrations or preference pairs. The skill is part of the wshobson fine-tuning sequence. It assumes the routing step has already decided that reinforcement learning is the right method, and it points to the separate skills for demonstration-based and preference-based training. The skill covers when RL is the right tool, a reference recipe for running GRPO, a mandatory gate that inspects the reward function before a run, and how to choose a GRPO variant when the base recipe behaves badly. The reward inspection step matters because a model will optimize whatever the reward measures, including shortcuts that do not reflect the real goal. It is installed with the skills CLI from the wshobson/agents repository. Its skills.sh page showed about 2.4K installs. The security audits showed Gen Agent Trust Hub with a warning, Socket passing and Snyk passing at the time of the scan.