Question about the code released

[Rain-Ju]

I'm having two problems trying your code.

1. I uploaded something about two minutes of video I took and asked the same question (how many chairs, tables, etc. in the room) and got some inaccurate answer, so I'm guessing what's causing this?
2. I tried to modify the prompt words inside the inference.py and realized that no matter what i ask, like what objects are in the room and expect him to give a detailed reasoning process. However the answer I get is still only 8 tokens and only numeric answers, what is the reason for this?

Thanks！

[diankun-wu]

Hi, the SFT checkpoints are fine-tuned on QA pairs with a certain format, so their outputs generally follow that predefined structure (for example, 8 tokens as you mentioned). Models that undergo further training with cold start and GRPO can first perform reasoning before providing an answer.

[Rain-Ju]

Thank you！So the models you mentioned that undergo further training with cold start and GRPO are released now? How to use?

[cwchenwang]

@diankun-wu when the model trained with GRPO can be provided?

[danielfaster]

Thanks a lot for this amazing repository, @diankun-wu! It’s a great contribution to the field of Spatial MLLMs.

I was wondering if there is a timeline for the release of the GRPO training code? It would be a fantastic addition to the current codebase.

Thank you for all your hard work!

[diankun-wu]

Thank for your interest!

At the moment, we don’t have a concrete release date to share, as our team is currently tied up with other work. That said, releasing the GRPO training code (and checkpoints) is on our roadmap.

I’ll keep this issue open and will update here once the GRPO training code and corresponding models are available.

Thanks again for your interest and patience!