Question: Which benchmark checkpoints and evaluation code are currently open-sourced?

[Ethan-Chen-plus]

Hi OpenBMB team,

Thanks for releasing MiniCPM-RobotManip. The README reports results on LIBERO (97.5), CALVIN ABC→D (4.1), RoboTwin2 easy/hard (91.3/91.6), and RMBench (53.3).

Could you please clarify:

1. Which checkpoint was used for each of the reported benchmark results?
2. Besides openbmb/MiniCPM-RobotManip, are any benchmark-specific, intermediate, or final training checkpoints already available?
3. Are the complete benchmark evaluation scripts, adapters, task configurations, action/state preprocessing, and evaluation logs available or planned for release?
4. If some checkpoints or evaluation code are not public yet, is there an estimated release timeline?

We can run the released checkpoint locally, but without the official benchmark adapters and evaluation details it is difficult to reproduce the reported numbers exactly.

Thank you!

[wzyaxyy]

Hi, thank you for your interest in our work.
The released weights are the general-purpose final weights and support multiple evaluation benchmarks.
Besides that, The evaluation code you asked about will be open-sourced soon—please stay tuned.