Question about datasets (test/validation split) and GFlowNet fine-tuning convergence

[xxiqiao]

Hi, thank you for your great work and for sharing the code! I have two questions regarding the paper and implementation:

1. Datasets (test/validation split): Could you clarify whether the paper used a separate validation set or test set during training? If so, how were they constructed?
2. GFlowNet fine-tuning convergence: In the fine-tuning stage with GFlowNet, how do you determine when training has converged?

Thanks again for your time, and I really appreciate your help!

[seanie12]

Sorry for the late reply.

1. There is no explicit test/validation dataset. After training, a prompt generator samples prompts and evaluate how many prompts can actually elicit harmful responses from target language models.

2. We monitor reward curves. If the reward plateau, we consider the model has converged.