More detials about sft on navsim

[JacheSha]

Dear Authors,
Thank you for open-sourcing DriveMA. The action-centric pretraining, meta-action SFT, and turn-level reinforcement learning designs presented in the paper have been very inspiring for my research.
I have recently been conducting some experiments on SFT reproduction. May I ask what the overall PDMS scores and component metrics of DriveMA-2B and DriveMA-4B are on NAVSIM after SFT but before reinforcement learning? These results would provide a valuable reference for my reproduction experiments.
Thank you for your time and excellent work. I look forward to your reply.

[wczheng]

Dear Zhiqi,

Thank you very much for your interest in our DriveMA work and for your kind recognition of our pretraining, SFT and RL designs. We are glad that our paper can bring inspiration to your research on SFT reproduction.

We sincerely apologize that some of our experimental archive files are temporarily missing at present, so we can only provide the NAVSIM metrics of DriveMA-2B after SFT and before reinforcement learning for your reference right now. We will sort out and share the complete results of DriveMA-4B with you as soon as we recover the missing files later.

The full PDMS score and all component metrics of DriveMA-2B are listed in the table below:

Model	PDMS	Collision	Area	Direction	Progress	TTC	Comfort
DriveMA-2B	86.55	98.57	94.55	98.05	80.83	94.44	99.98

If you have any other questions about reproducing the SFT of DriveMA-2B, please feel free to let us know. We will reply to you promptly.

Thank you for your understanding and support.

Best regards,
Weicheng Zheng

[JacheSha]

@wczheng ok thank you a lot ！

[wczheng]

You're very welcome!
Hope these metrics can facilitate your reproduction work. If our project brings you some help, a star would be highly appreciated!

[JacheSha]

Hi,

Thank you for sharing these details in the paper. I have a few questions regarding the RL setup and would really appreciate it if you could clarify.

In the paper, it's mentioned that the KL penalty coefficient is set to β = 0.4 with a sampling temperature of 1.0. However, I noticed that in the provided code, the value appears to be β = 0.04, along with other parameters like:

text
--beta 0.04
--epsilon 3e-4
--epsilon_high 4e-4
--steps_per_generation 1
--importance_sampling_level sequence
--num_train_epochs 20
I assume the above configuration corresponds to the WOD-E2E RL experiments. Could I kindly ask:

Is there a specific reason for the discrepancy in the β value between the paper and the code?
For the NAVSIM RL experiments, could you share the training epochs, training steps, and the value of the KL coefficient β you used?
Thank you very much for your time and help — I look forward to hearing from you.