It seems that there is only one script for generating the training dataset:
bash scripts/run_nuplan_preprocessing.sh
I have a few questions:
1.Are both the SFT data and RL data generated by this unified preprocessing script?
2.What are the key differences in how SFT data and RL data are constructed?
3.How do we determine the appropriate SFT-to-RL data ratio when training on the full dataset, such as WOD-E2E and nuPlan?
It seems that there is only one script for generating the training dataset:
bash scripts/run_nuplan_preprocessing.sh
I have a few questions:
1.Are both the SFT data and RL data generated by this unified preprocessing script?
2.What are the key differences in how SFT data and RL data are constructed?
3.How do we determine the appropriate SFT-to-RL data ratio when training on the full dataset, such as WOD-E2E and nuPlan?