Skip to content

Hugging Face Space Deployment

Prerequisites

  • A Hugging Face account with write access
  • HF_TOKEN set in your environment
  • Dockerfile-based Space (this repo uses Docker)

Deploy

export HF_TOKEN="<your_hf_token>"
python scripts/deploy_hf_space.py \
  --repo Mo-alaa/agentic-safety-eval \
  --token "$HF_TOKEN"

Add --no-create if the Space already exists and you are pushing an update.

What Gets Deployed

  • Full repository codebase (excluding large dev artifacts)
  • FastAPI backend (server/) and built frontend assets (frontend/dist)
  • Dockerfile — Space runs the container on port 7860

Required Space Secrets

Configure these in the Hugging Face Space settings after deployment:

Secret Purpose
GENAI_STUDIO_API_KEY Primary inference API
OPENAI_API_KEY OpenAI model access
GEMINI_API_KEY Gemini model access
ANTHROPIC_API_KEY Anthropic model access
OLLAMA_CLOUD_API_KEY Ollama cloud endpoint
WANDB_API_KEY Experiment tracking
HF_RESULTS_DATASET Set to Mo-alaa/agentic-safety-results to auto-mirror results
HF_TOKEN Readable token for private dataset access

When HF_RESULTS_DATASET is set, the backend auto-mirrors all files under results/ from the dataset into local Space storage on first access to /api/results routes.

Local Validation

Test the server before deploying:

python -m uvicorn server.main:app --host 0.0.0.0 --port 7860

Or with Docker:

docker build -t agentic-safety .
docker run --rm -p 7860:7860 agentic-safety

API Routes

The deployed Space exposes:

  • /api/results — list available result directories
  • /api/results/{id} — fetch a specific result's summary and records
  • /api/jobs — job management for experiment launches

→ Results browsing via HF Dataset