OpenVLA-7B LIBERO-Spatial with early-layer VQ
This is the assembled, merged full-model checkpoint from a 50K-step LIBERO-Spatial run based on openvla/openvla-7b-finetuned-libero-spatial. It adds the early visual vector-quantization module used in this project and does not enable FiLM. The VQ layer is inserted at visual block index 3 with a 1024-entry full-dimensional codebook. The uploaded weights include the VQ codebooks and do not require loading a separate LoRA adapter.
Training configuration: 4 H100 GPUs, per-GPU batch size 16, seed 7, 50,000 optimizer steps.
Loading
The repository includes the custom Transformers remote-code files and processor assets. Use trust_remote_code=True when loading:
from transformers import AutoModelForVision2Seq, AutoProcessor
repo_id = "Jiahao-Wang/openvla-7b-libero-spatial-vq-b3-k1024-b64-r32-50k-seed7-v1"
processor = AutoProcessor.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForVision2Seq.from_pretrained(
repo_id, trust_remote_code=True, torch_dtype="bfloat16", device_map="auto"
)
The action normalization statistics for libero_spatial_no_noops are stored in config.json.