Hi, thanks for the great work!
I have a question regarding the training dataset.
The paper mentions that the model is trained on a 1M dataset following PLUTO.
I would like to know whether the released checkpoint is trained using:
the original PLUTO training_scenarios_1M scenario filter without any modifications, or
a custom scenario split / filtering strategy.
Specifically,
Is the 1M dataset generated directly from PLUTO's original code?
Did you modify the scenario filter (e.g., scenario types, log selection, sampling strategy, or random seed)?
Is the released checkpoint reproducible using the public preprocessing pipeline described in the repository?
I ask because I reproduced the preprocessing pipeline but limited the maximum number of samples per scenario type (3000 each), and my trained model performs noticeably worse than the released checkpoint. I wonder whether the difference comes from the dataset construction.
Thanks!
Hi, thanks for the great work!
I have a question regarding the training dataset.
The paper mentions that the model is trained on a 1M dataset following PLUTO.
I would like to know whether the released checkpoint is trained using:
the original PLUTO training_scenarios_1M scenario filter without any modifications, or
a custom scenario split / filtering strategy.
Specifically,
Is the 1M dataset generated directly from PLUTO's original code?
Did you modify the scenario filter (e.g., scenario types, log selection, sampling strategy, or random seed)?
Is the released checkpoint reproducible using the public preprocessing pipeline described in the repository?
I ask because I reproduced the preprocessing pipeline but limited the maximum number of samples per scenario type (3000 each), and my trained model performs noticeably worse than the released checkpoint. I wonder whether the difference comes from the dataset construction.
Thanks!