Evaluation Dataset Missing, Training Instability, and Unclear Hyperparameter Settings

[WinterPenguin2]

Hello,

I encountered several issues while trying to reproduce the results:

1. Evaluation Dataset
   The dataset required for evaluation is not provided.
   The evaluation path is hard-coded to the author’s local desktop, making evaluation impossible after training.
   Could you please provide the dataset or a valid path?

2. Training Instability

When running on the experiments, the training loss consistently drops to 0.
In GRPO, the Advantage term always becomes 0 (huggingface/open-r1#239 (comment)), but training is still possible since the KL divergence term is non-zero.

However, in Visual RFT, the KL divergence coefficient is set to 0, which makes the training dynamics unclear.
I am unsure if this is intended or a mistake.

Could you please share the missing hyperparameters or supplementary material that includes detailed training logs? This would help in reproducing and validating the reported results.

Thank you!