OpenVLA

SkillDev tools

Use when working on OpenVLA fine-tuning/serving/evaluation workbench stages, the OpenVLA-OFT LoRA recipe, or the npa workbench openvla CLI.

Available today. Use it from your connected AI after setup.

Add ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.

Then ask your AI: use the OpenVLA skill

What this skill tells your AI

The instructions your AI receives, as published by nebius/nebius-physical-ai in skills/tools/openvla/SKILL.md and read by ahel’s review.

OpenVLA is the vision-language-action tool for fine-tuning, serving, and evaluating OpenVLA policies (OpenVLA-OFT recipe) in npa.

The three core stages (train, serve, eval) are currently stubs: they expose the intended argument signatures but exit with a clear "not yet implemented" message until the OpenVLA-OFT fine-tuning pipeline lands (tracking issue nebius/nebius-physical-ai#500). The base model (default openvla/openvla-7b) is resolved at runtime through the HF Hub cache via npa.workbench.model_access — weights are never bundled in the repo or image.

Interfaces

  • CLI: npa workbench openvla <train|serve|eval> --help
  • Python SDK: npa.sdk.workbench.openvla (train, serve, eval)
  • Workflow module: npa.workflows.byof.openvla_pipeline (argument validation, upstream argv planning, argparse entrypoint)

Conventions

  • The CLI lives at npa.cli.workbench.openvla (multi-command package form for new tools); the SDK surface lives at npa.sdk.workbench.openvla.
  • The train stage is the headline openvla/train three-tier contract (CLI <-> SDK <-> workflows/testing/openvla-train.yaml); serve and eval remain public-reusable sibling toolRefs.

Signals

GitHub stars
30
Forks
16
Last commit
Sep 2026
Advanced
Catalog kind
skill
Key
openvla
Source
github.com/nebius/nebius-physical-ai