Hugging Face Skills: Teaching Claude Code and Cursor to Train AI Models Directly

Hugging Face has released an official Skills repository that gives Claude Code, Cursor, OpenAI Codex, Gemini CLI, and other major coding agents full AI/ML pipeline capabilities — already 5k+ stars and climbing fast. No more wrestling with training environments; just tell your AI what you need and it handles the entire chain from data prep to model deployment.

Currently offers 9 skills covering core ML workflows: model training (SFT/DPO/GRPO/reward modeling with automatic hardware selection and cost estimation), dataset management (init/update/SQL queries), model evaluation (auto-evaluate after training, write scores to model cards), paper publishing (index arXiv papers on Hub, link to models and datasets), and cloud jobs (run Python scripts on HF infrastructure, zero server management). Also includes Gradio UI building and Trackio experiment tracking.

Uses the Agent Skill standard format (one folder + SKILL.md per skill), compatible with all major coding agents. After installation, invoke via natural language — e.g., "Use the HF trainer skill to estimate GPU memory for a 70B model run" — and the agent auto-loads instructions and scripts.

Background

Hugging Face has released an official Skills repository enabling Claude Code, Cursor, OpenAI Codex, Gemini CLI, and other major coding agents to gain full AI/ML workflow capabilities. Already 5k+ stars — clear evidence of strong demand for "ML pipeline via natural language."

Previously, training models meant setting up environments, writing training scripts, configuring hardware, and manually running evaluations. Now install a Skill, tell your AI what you need, done.

9 Skills Explained

1. Model Training (hugging-face-model-trainer)

  • SFT, DPO, GRPO, reward modeling via TRL on HF Jobs infrastructure
  • **Automatic hardware selection and cost estimation**: specify model size, get GPU config and pricing
  • GGUF conversion for local deployment, Trackio monitoring, Hub persistence

2. Dataset Management (hugging-face-datasets)

  • Initialize repos, define configs/system prompts, stream row updates
  • **SQL-based querying and transformation**

3. Model Evaluation (hugging-face-evaluation)

  • Extract eval tables from README, import scores from Artificial Analysis API
  • Run custom evaluations with vLLM/lighteval
  • **Auto-write scores to model cards**

4. Paper Publishing (hugging-face-paper-publisher)

  • Publish arXiv papers to HF Hub, auto-link to models/datasets, claim authorship

5. Cloud Jobs (hugging-face-jobs)

  • Execute Python scripts on HF infrastructure, manage scheduled jobs, zero server management

6. CLI Operations (hugging-face-cli)

Download models/datasets, upload files, manage repos, run cloud compute

7. Gradio UI (gradio)

Build web UIs and demos — components, event listeners, layouts, chatbots

8. Experiment Tracking (hugging-face-trackio)

Python API for logging metrics, CLI queries, real-time dashboards synced to HF Spaces

9. Tool Builder (hugging-face-tool-builder)

Build reusable HF API scripts, chain API calls, automate repetitive tasks

Compatibility

Agent Skill standard format (folder + SKILL.md + YAML frontmatter):

  • **Claude Code**: `/plugin install <skill>@huggingface/skills`
  • **OpenAI Codex**: Copy to `.agents/skills/`, auto-discovered
  • **Gemini CLI**: `gemini extensions install` from GitHub URL
  • **Cursor**: Via `.cursor-plugin/plugin.json` and `.mcp.json`

Fallback: `agents/AGENTS.md` for agents without skill support.

Usage Examples

  • "Estimate GPU memory for a 70B model run" → full plan + config code
  • "Run evaluation on latest checkpoint" → auto-execute and write results
  • "Create few-shot classification templates" → initialize dataset repo
  • "Index my arXiv paper and link to my model" → auto-index and link

Significance

This is standardized infrastructure for AI-assisted ML. HF as the core ML ecosystem platform building Skills equals providing "ML capability plugins" for the entire AI Agent ecosystem.

License: MIT | Stars: 5k+ | Maintained by: Hugging Face