Inquiry about evaluation

Thank you for the good work!

I have some questions regarding the evaluation of MOSS-VL-Realtime on streaming benchmarks, particularly on the proactive tasks of StreamingBench and OVO-Bench. Since the two benchmarks both evaluate a model's proactivity in a pseudo-streaming manner (OVO-Bench queries the model at several preset time points, while StreamingBench polls the model every second and terminates upon the first positive response. Both of them do not actually let the model make the decision to response or remain silent by itself.), so I am wondering if you have followed the established evaluation protocals of these two benchmarks, or make some alterations to align them with real streaming settings.

If possible, could you please open-source your evaluation code? That would be a great help to me! Thank you again for your impressive work!

Start by reviewing the evaluation protocols for StreamingBench and OVO-Bench described in the issue, then read the discussion for any maintainer response. Done would require a clarified account of the protocols used and, if agreed, the evaluation code being made available.