Hello, and thank you for the great work!
I have a question regarding the training configuration. In the README, the provided training command includes the following setting:
--max_train_steps 10000
However, the paper mentions that the dataset contains around 80k samples. With a batch size of 1, training for only 10k steps seems significantly smaller than what would typically be required to fully utilize the dataset or achieve the results reported in the paper.
My questions are:
What is the actual number of training steps used in the paper?
Is the 10000 in the README intended only as an example, rather than the configuration used for the final model?
If possible, could you share the actual training schedule (total steps or epochs) used in the experiments?
Thank you again for releasing this excellent work. I appreciate your help!
Hello, and thank you for the great work!
I have a question regarding the training configuration. In the README, the provided training command includes the following setting:
--max_train_steps 10000
However, the paper mentions that the dataset contains around 80k samples. With a batch size of 1, training for only 10k steps seems significantly smaller than what would typically be required to fully utilize the dataset or achieve the results reported in the paper.
My questions are:
What is the actual number of training steps used in the paper?
Is the 10000 in the README intended only as an example, rather than the configuration used for the final model?
If possible, could you share the actual training schedule (total steps or epochs) used in the experiments?
Thank you again for releasing this excellent work. I appreciate your help!