what does this mean? Clone the source code of Open-R1 from git clone https://github.com/huggingface/open-r1. Merge and replace files between ours and Open-R1's (including setup.py).

[pengzhangzhi]

Hi,

great work! I'm trying to reproduce the results. But this doesnt make much sense to me Clone the source code of Open-R1 from git clone https://github.com/huggingface/open-r1. Merge and replace files between ours and Open-R1's (including setup.py). As in what do you mean by replace files between ours and Open-R1's (including setup.py). I'm totally confused. Could you be more specific?

[pengzhangzhi]

recipes/config_demo_code.yaml this config is also missing. I wonder if the authors can provide a working code?

[summmeer]

Sorry, the training config is recipes/config_coupled_code.yaml. We will update this.
We work on open-r1 code base, with some modifications. Currently we only list those files that modified by us. You can use the source code and do some merge and replacement to acheive the same function of our working code space.

[pengzhangzhi]

Hi, thanks!

Im not positive that the config recipes/config_coupled_code.yaml can run, all the arguments are just dummies.

# Model arguments
model_name_or_path: /path-to-model
torch_dtype: bfloat16
trust_remote_code: true
# Data training arguments
dataset_name: /path-to-dataset

Has the team — or anyone else — successfully reproduced the reported results using this codebase? I highly suggest the authors test and run ur code before releasing it. It would be really helpful to have a working example or a more complete configuration to get started.

I really enjoyed the paper and appreciate the authors’ transparency in sharing their research. That said, if the goal is to open-source the work, I hope the codebase can be made more reproducible and complete — so that it’s not just open in appearance, but truly usable by the community.

[summmeer]

Hi, you can easily replace these parameters into your own ones. For make the code clearer, we update the config and inference examples. For coupled-GRPO training, we provide the core implementation and list all files that modified by us. You need to read the source code of open-r1. This current version is released after the consideration of legal issues.