Osunz/openvla_FT_libero

OpenVLA Fine-Tuned on LIBERO Spatial

This repository hosts my fine-tuned OpenVLA checkpoint trained on the modified LIBERO Spatial dataset variant libero_spatial_no_noops.

Model Summary

• Base model: openvla-7b
• Fine-tuning method: LoRA
• Training steps: 50,000
• Final evaluation suite: libero_spatial
• Final score: 40/50 = 80.0%

Dataset

This model was trained on:

• Dataset name: libero_spatial_no_noops
• Dataset format: RLDS / TFRecord
• Dataset download: https://huggingface.co/datasets/openvla/modified_libero_rlds/tree/main/libero_spatial_no_noops/1.0.0

Place the dataset under:

datasets/modified_libero_rlds/libero_spatial_no_noops/1.0.0

Checkpoint Contents

This Hugging Face model repository contains the final merged fine-tuned checkpoint exported from the run directory:

runs/openvla-7b+libero_spatial_no_noops+b16+lr-0.0005+lora-r32+dropout-0.0--image_aug

Upstream Project

This model is based on the upstream OpenVLA project:

https://github.com/openvla/openvla

License

• Repository code in upstream OpenVLA is released under the MIT License
• Pretrained OpenVLA models may also be subject to the Llama Community License

Llama license:

https://ai.meta.com/llama/license/