Request for the exact Stage 2 GRPO subset after low-variance filtering

[qure-shivam-shukla]

Dear Authors,
Thank you for open sourcing CheXOne, this is a really impressive and carefully executed piece of work, and I especially appreciated how clearly the training stages were described in the paper.

I wanted to ask whether you could directly share the exact Stage 2 GRPO training subset after low-variance filtering. It would be very helpful to have the final subset itself (the exact file used), rather than just the preprocessing script.

Thanks again for releasing this work.

[YBZh]

Thank you for your attention.

Regarding the Stage 2 GRPO training subset after low-variance filtering, the filtered subset is highly model-dependent. In particular, it depends on the model checkpoint, reward function used during filtering. Therefore, the exact subset used in our run may not be optimal or fully reproducible for a different model or training setup.

For this reason, we recommend processing the dataset with your own model and filtering configuration when developing or adapting your model. This should produce a subset that is better aligned with your model’s response distribution and training dynamics.

[qure-shivam-shukla]

Hi, thank you for the clarification.

That makes sense in the general case. In our case, though, our goal is to follow the same Stage 2 GRPO setup used for CheXOne as closely as possible, rather than re-filtering the data with a different model or reward configuration.

Because of that, we were hoping to directly use the exact filtered subset from your CheXOne GRPO run. This would help us avoid the extra sampling/filtering step and the associated compute cost, while staying as close as possible to the training configuration described in the paper.

So, please share the final filtered subset itself used in your run. That would be very helpful.

Thanks again for your time and for releasing this work.