Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.cyberwave.com/llms.txt

Use this file to discover all available pages before exploring further.

Overview

OpenVLA-OFT is a LoRA-based fine-tuning framework for Vision-Language-Action models built on the openvla-7b base. Cyberwave supports training and deploying OpenVLA-OFT models on your custom datasets using the openvla (RLDS) dataset format.
OpenVLA-OFT uses the openvla (RLDS) dataset format for training, which differs from the LeRobot v3 format used by SmolVLA. The platform handles format selection automatically based on the model type.

Model Selection

When creating an ML model on Cyberwave, select OpenVLA-OFT as the model type to automatically route training and inference jobs to a cloud node with the openvla-oft-policy profile.

Training

Training jobs are dispatched to a Cyberwave Cloud Node. The cloud node:
  1. Downloads the dataset in RLDS format (format=openvla)
  2. Downloads the base model weights (openvla/openvla-7b by default)
  3. Runs the LoRA fine-tuning loop, routing all metrics to the Cyberwave dashboard
  4. Compresses and uploads the final checkpoint to the platform

Key training parameters

ParameterDescriptionDefault
max_stepsNumber of training gradient steps50 000
batch_sizePer-GPU batch size (auto-scaled to camera count if omitted)auto
lora_rankLoRA adapter rank32
use_l1_regressionUse L1 regression action head instead of cross-entropytrue
use_proprioInclude proprioceptive state in observationstrue
num_images_in_inputNumber of camera feeds used during training (auto-detected if omitted)auto

Inference

Trained checkpoints can be deployed directly from the Cyberwave platform. The cloud node loads the fine-tuned LoRA weights and streams predictions back to your robot via MQTT.

Dataset format

OpenVLA-OFT requires datasets in the RLDS (TensorFlow Datasets) format. When you upload a dataset through the Cyberwave platform, it is automatically converted and made available for download in the openvla format.