Repository navigation
Releasing SFT checkpoint and on OPD/SOD training details #20
Description
Hi authors, thanks so much for open-sourcing this great work!!
I am trying to reproduce the results using the released SFT data. I first trained an SFT checkpoint from scratch using the provided data and then ran both OPD and SOD on top of it. I noticed that my reproduced SFT checkpoint does not match the scores reported in the paper, especially on the LiveCodeBench. I ran OPD and SOD models based on the SFT checkpoint, and the numbers also couldn't match those reported in the paper.
So I was wondering:
-
Would it be possible to release the SFT checkpoint used in the paper?
It would be very helpful for the community to study on the performance in the SOD/OPD stage itself. -
How many training steps were used for the SOD and OPD experiments, respectively?
In another GitHub issue, the authors mentioned:
"For the 4B→1.7B SOD experiment, we observed that the model converges quite early in training."
In my experiments, after training for around 150–200 steps, the performance still remains low, and the model even starts to show signs of degradation instead of convergence (not sure if this is because of the SFT checkpoint I was using).
Any additional details about the training setup would be greatly appreciated!
Thanks a lot for your help and thanks again for sharing details of your great work!
Activity
Qijia-He commentedon Jul 6, 2026
I’m running into a similar issue and it would be very helpful if the authors could release the SFT checkpoint used in the paper.
YoungZ365 commentedon Jul 11, 2026
Thank you very much for your interest in our work and for the detailed reproduction effort!
Q1. Would it be possible to release the SFT checkpoint used in the paper?
Yes. We are currently preparing the SFT checkpoint used in the paper and plan to release it soon.
Q2. How many training steps were used for the SOD and OPD experiments?
Please refer to Issue #19, where we provide more details regarding the training steps and convergence behavior. In our experiments, the model typically converges relatively early under the intended training setup.
Q3. After training for around 150–200 steps, the performance still remains low...
Please also refer to Issue #22. The behavior you described (remaining at a low performance level or even degrading after 150–200 steps) is usually not what we observe under a correctly configured training environment. As discussed there, one common cause is issues with the sandbox/tool execution environment (e.g., failed tool execution, insufficient concurrency, or unstable sandbox services), which can significantly affect training stability.
Thanks again for your support and for taking the time to reproduce our work!