---

name: hugging-face description: This skill should be used when the user asks about "Hugging Face", "HF Hub", "transformers", "model hub", or needs guidance on which Hugging Face capability to use. Acts as an entry-point that routes to specialized HF skills (cli, jobs, datasets, evaluation, model-trainer, paper-publisher, trackio, tool-builder) based on the task. Use for authentication setup, quick operations, and choosing the right specialized skill. version: 1.0.0 license: MIT

Hugging Face Integration Hub

Entry-point orchestrator for Hugging Face capabilities. Routes to specialized skills based on task.

---

Quick Start: Authentication

Before using any Hugging Face functionality, configure authentication:

Copy# Option 1: Environment variable (recommended)
export HF_TOKEN="hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

# Option 2: CLI login (interactive)
huggingface-cli login

# Option 3: Python
from huggingface_hub import login
login(token="hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxx")

Verify authentication:

Copyhuggingface-cli whoami

Use scripts/auth_setup.py for guided authentication configuration.

---

Skill Routing Table

Choose the right specialized skill based on task:

---

Decision Tree

"I need to download or upload files"

→ Use hugging-face-cli

Copyhf download meta-llama/Llama-3.1-8B
hf upload my-model ./model --repo-type model

"I need to run computation (training, inference, batch processing)"

→ Use hugging-face-jobs

• GPU training jobs
• Batch inference
• Data processing pipelines

"I need to create or query a dataset"

→ Use hugging-face-datasets

• Create dataset repos
• Add rows via streaming
• SQL-based transformations

"I need to benchmark or evaluate a model"

→ Use hugging-face-evaluation

• Run lighteval/vLLM evaluations
• Add results to model cards
• Compare against leaderboards

"I need to fine-tune a model"

→ Use hugging-face-model-trainer

• SFT (Supervised Fine-Tuning)
• DPO (Direct Preference Optimization)
• GRPO (Group Relative Policy Optimization)
• GGUF conversion

"I need to publish a research paper"

→ Use hugging-face-paper-publisher

• Create paper pages
• Link to models/datasets
• Claim authorship

"I need to track training metrics"

→ Use hugging-face-trackio

• Log metrics during training
• Create visualization dashboards
• Sync to HF Spaces

"I need to build custom automation"

→ Use hugging-face-tool-builder

• API scripting patterns
• Reusable tooling
• Complex workflows

---

Common Quick Operations

Search for Models

Copyfrom huggingface_hub import HfApi

api = HfApi()

# Search by task
models = api.list_models(
    task="text-generation",
    sort="downloads",
    direction=-1,
    limit=10
)

for model in models:
    print(f"{model.modelId}: {model.downloads:,} downloads")

Download a Model

Copy# Using CLI
hf download meta-llama/Llama-3.1-8B --local-dir ./llama

# Using Python
from huggingface_hub import snapshot_download
snapshot_download("meta-llama/Llama-3.1-8B", local_dir="./llama")

Upload Files

Copy# Upload single file
hf upload my-username/my-model ./model.safetensors

# Upload directory
hf upload my-username/my-model ./model_dir --repo-type model

# Create repo and upload
hf repo create my-model --type model
hf upload my-username/my-model ./model_dir

Check Model Info

Copyfrom huggingface_hub import model_info

info = model_info("meta-llama/Llama-3.1-8B")
print(f"Downloads: {info.downloads:,}")
print(f"Likes: {info.likes}")
print(f"Tags: {info.tags}")

Use scripts/quick_search.py for fast model/dataset lookup.

---

Environment Variables

---

API Endpoints

---

Inference API (Quick)

For quick inference without downloading models:

Copyimport requests

API_URL = "https://api-inference.huggingface.co/models/gpt2"
headers = {"Authorization": f"Bearer {HF_TOKEN}"}

response = requests.post(
    API_URL,
    headers=headers,
    json={"inputs": "Hello, I'm a language model"}
)
print(response.json())

For dedicated inference endpoints and production workloads, use hugging-face-jobs.

---

Bundled Resources

• scripts/auth_setup.py - Guided authentication configuration
• scripts/quick_search.py - Fast model/dataset search
• references/skill-routing.md - Detailed decision tree
• references/quick-reference.md - Common operations cheat sheet

---

Specialized Skill Summaries

hugging-face-cli

File operations, downloads, uploads, repo management via the hf CLI tool.

hugging-face-jobs

Run workloads on HF infrastructure: UV scripts, Docker jobs, GPU selection, cost estimation.

hugging-face-datasets

Create datasets, define configs, stream rows, SQL transformations, dataset querying.

hugging-face-evaluation

Model evaluation, benchmark running, metric extraction, model card updates.

hugging-face-model-trainer

Fine-tuning with TRL (SFT, DPO, GRPO), GGUF conversion, Trackio integration.

hugging-face-paper-publisher

Research paper pages, model/dataset linking, authorship claims, markdown articles.

hugging-face-trackio

Experiment tracking, metric logging (Python API), metric retrieval (CLI), HF Space sync.

hugging-face-tool-builder

Build reusable scripts for HF API operations, automation patterns, workflow templates.