Actor and critic checkpoints obtained by RL training on top of Qwen3-4B-Instruct-2507.

Checkpoints under directories whose names end in _hf (for example ckpts/<run>_hf/iter_XXXXXXX) are in Hugging Face format and can be loaded directly with transformers. All other checkpoints are in Megatron torch_dist format.

Each checkpoint directory <dir> has an MD5 checksum list <dir>.md5 next to it; check it with md5sum -c ../<dir name>.md5 from inside the directory.

License: see LICENSE.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for oooyc/rppo-ckpts

Base model

Qwen/Qwen3-4B-Instruct-2507

Finetuned

(2608)

this model