nokaikai/spatialvla_cocacola_lora

SpatialVLA LoRA fine-tuned weights on the Cola dataset

SpatialVLA LoRA fine-tuned weights on the Cola dataset (Franka + joint-space actions).

Model

• Base: SpatialVLA-4B-224-PT (LoRA: r=32, alpha=32, all attention + MLP layers)
• Data: nokaikai/cola_lerobot_v2 (7 joints + 1 gripper, absolute joint targets)
• Training: lr=1e-4, bs=4, 1 GPU, 10 epochs, linear scheduler, checkpoint-68390
• Use: Load for inference; input image + current 8-dim state + prompt -> output 8-dim action (7 joints + 1 gripper)

Usage

pip install huggingface_hub
huggingface-cli download nokaikai/spatialvla_cocacola_lora --local-dir ./spatialvla_cocacola_lora

Loading from code:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("pretrained/spatialvla-4b-224-pt")
model = PeftModel.from_pretrained(base_model, "./spatialvla_cocacola_lora")

Layout

checkpoint-68390/
  adapter_config.json          # LoRA config (r=32, alpha=32)
  adapter_model.safetensors    # LoRA weights
  processing_spatialvla.py     # SpatialVLA processor
  processor_config.json
  action_tokenizer.py
  tokenizer.json / tokenizer_config.json

License & Credits

Weights trained with SpatialVLA and Cola data; for research/personal use. Cola dataset: nokaikai/cola_lerobot_v2 on Hugging Face.