Hugging Face

SkillDatabases & data

Inspect and manage robotics datasets, models, Jobs, and Spaces on Hugging Face.

Then ask your AI: use the Hugging Face skill

Details

Instructions available. Your AI can read the instructions. Execution depends on the setup they require.

Add it as a step in a pipeline you build on Ahel, or connect Ahel to Claude, ChatGPT, Cursor, Claude Code or Codex and ask your AI to use it.

What this skill tells your AI

The instructions your AI receives, as published by robium-ai/robium in skills/huggingface/SKILL.md and read by Ahel’s review.

The Hub is an artifact boundary. Know exactly what is crossing it, under which identity and revision, before transferring data or changing remote state.

Inspect first

  • Check the installed hf CLI and the relevant command group's live help. Commands and hosted services evolve faster than this skill.
  • Inspect the repository card, file tree, revision, license, access state, and transfer size before downloading or planning around an artifact.
  • For robot data, also verify embodiment, task, observation/action schema, cameras, timing, episode structure, and format version.
  • Pin an immutable revision when the artifact feeds a reproducible build, fixture, evaluation, or training run.

Treat remote changes as real changes

  • Use an existing authenticated identity or a secure browser/device login. Do not place tokens in prompts, source, shell history, or logs.
  • Before creating or uploading, establish the destination owner, repository type, visibility, local source, changed paths, and whether deletion is involved.
  • Starting Jobs is paid remote compute. Establish the current hardware, command, timeout, outputs, secrets, and cost exposure before submission.
  • Inspect a Space's metadata, build logs, and runtime logs before restarting or changing hardware.
  • A Space is an optional hosted presentation layer, not a requirement for using Hub datasets or models. Most Docker Spaces can also be run locally; keep the application portable when self-hosting or another service is plausible.
  • After a mutation, inspect the remote result rather than trusting the local command's intent.

Go deeper only when needed

  • For current CLI shapes, Dataset Viewer endpoints, and mutation preflights, read references/hub-operations.md.
  • LeRobot owns dataset structure, processors, training, and evaluation; this skill owns the Hub boundary around those artifacts.
  • Data owns source selection. Use integration for service packaging, network, health, and self-hosting boundaries; use the relevant deployment skill for provider mechanics when a Space is not the chosen runtime.
  • Use the current Hugging Face documentation and live CLI help as the source of truth for volatile flags and services.

Done

  • The intended artifact, revision, identity, destination, and remote result are all verified, with no credential exposed and no unapproved paid or destructive action.

Signals

GitHub stars
88
Forks
1
Last commit
Oct 2026
Advanced
Item type
skill
Key
huggingface-robium-ai
Source
github.com/robium-ai/robium