huggingface_workshop

This repo contains the materials for a hands-on workshop on Hugging Face datasets and models.

Workshop Agenda

1. Navigate the Hugging Face ecosystem

   • Explore Spaces, tasks, datasets, and models on the Hub
   • Read dataset and model cards
   • Understand why Hugging Face formats are well suited for open source
     • Parquet for datasets
     • Safetensors for model weights
   • Work with the datasets and transformers libraries

2. Use Hugging Face MCP with Claude Code or Codex

   • Access Claude through a Harvard billing account
   • Access Codex with your HUID
   • Explore MCP-powered workflows

3. Run models in minutes

   • Prepare the Restor dataset with datasets
   • Perform inference with transformers
   • Compare NVIDIA SegFormer and TCD-SegFormer on a test sample

4. Build an end-to-end training loop

   • Fine-tune NVIDIA SegFormer on the Restor dataset
   • Track evaluation metrics
   • Briefly explore SAM (Segment Anything Model) as a modern foundation model for segmentation

Harvard Setup

export ANTHROPIC_BEDROCK_BASE_URL=https://apis.huit.harvard.edu/ais-bedrock-llm/v2
export ANTHROPIC_API_KEY=$(cat api_key.txt)
export ANTHROPIC_SMALL_FAST_MODEL=us.anthropic.claude-opus-4-5-20251101-v1:0
export CLAUDE_CODE_SKIP_BEDROCK_AUTH=1
export CLAUDE_CODE_USE_BEDROCK=1

Materials

• Website
• Follow Along Notebook
• Follow Along Prefilled
• Notebook

Requirements

If you want to run the notebooks locally:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt