Framework and memory mode for the released RMBench checkpoint

[zoeyliu1999]

Hi authors, thank you for releasing the RMBench checkpoint!

The released RMBench/config.yaml describes the visual-anchors-only model as:

framework:
  name: QwenOFT

datasets:
  vla_data:
    keyframe_image_memory:
      enabled: false

memory_buffer:
  enable: false
  qwen_memory_injection:
    enabled: true
    mode: raw_anchors_only

However, the current public repository does not appear to register a QwenOFT framework.

In EventVLA/eventvla/model/memory_ablation.py, the only canonical mode is pure_image_keyframe_memory, which enables keyframe-image memory and therefore does not appear equivalent to the released RMBench visual-anchors-only config.

Could you clarify the exact public framework class and memory configuration that should be used to load RMBench/final_model/pytorch_model.pt without changing its intended visual-anchors-only inference behavior?

Thank you!