Notice

This is a private evaluation release. Authorized collaborators may inspect it for review, but no open-source license or redistribution permission is granted by this repository. Model weights, datasets, service credentials, and raw run artifacts are not included.

The implementation interoperates with Megatron-LM, Megatron-Bridge, verl, PyTorch, Ray, and related upstream projects. Their names, copyrights, and license terms remain with their respective authors. Before external release, the repository owner must add the applicable license for newly authored code and preserve all required upstream notices.

glm52/dapo_math_reward.py is a narrow adaptation of verl's DAPO math reward implementation. It retains the upstream ByteDance, EleutherAI, and Hugging Face copyright notices and is licensed under Apache License 2.0 as stated in that file. This repository notice does not replace or restrict that file-level license.