Inquiry about the "Capability-expansion datasets" SFT Dataset in InternVL3.5

[shimudong]

Thank you for your incredible work on the InternVL series, especially the latest InternVL3.5.

I've been studying your paper (arXiv:2508.18265) in detail, and I'm particularly fascinated by the significant improvements in spatial and embodied reasoning capabilities (ref VSI-Bench).

The paper mentions that the SFT stage for InternVL3.5 includes "Capability-expansion datasets" to endow the model with new skills, specifically citing "embodied interaction." This data appears to be a key factor behind the model's enhanced performance in this area.

Is this dataset is already included within the MMPR v1.2 dataset, or is it a distinct dataset used exclusively for SFT? If it's a distinct dataset, are there any plans to open-source this specific "Capability-expansion datasets" portion of the SFT training data?