StarVLA CALVIN ABC to D candidate
This repository stores a StarVLA checkpoint produced by WMH experiments.
Intended use
- Backbone: Qwen3-VL-4B-Instruct-Action base VLM.
- Action stack: StarVLA / GR00T-style action head, MoE95k initialization plus fresh LoRA continuation.
- Training data: CALVIN ABC and in-house ABC augmentations only.
- Evaluation target: CALVIN ABC to D.
Important paths at upload time
- Checkpoint source:
/inspire/qb-ilm2/project/26summer-camp-10/public/seven/starvla_calvin/members/WMH/runs/abc_moe95k_lora_aug_3h_bs96_0520_045012/checkpoints/steps_7000_pytorch_model.pt
- Run source:
/inspire/qb-ilm2/project/26summer-camp-10/public/seven/starvla_calvin/members/WMH/runs/abc_moe95k_lora_aug_3h_bs96_0520_045012
- n100 metrics source:
not uploaded by default
- n300 metrics source:
not uploaded by default
Do not treat this as an upstream pretrained action policy. It is a project checkpoint trained within the CALVIN ABC experiments.
