Using Huggingface for Specialized Models and Fine-tuning

This two-day, hands-on course introduces participants to the Hugging Face ecosystem, equipping them with practical skills to find, run, fine-tune, and deploy pre-trained models for real-world applications. Through guided exercises, attendees will explore NLP, vision, and audio models, learn the fundamentals of fine-tuning, and deploy their own models using Hugging Face Spaces. _ The course covers the full workflow—from navigating the Hugging Face Hub to preparing datasets, customizing models, and optimizing deployments—ensuring participants gain both conceptual understanding and practical coding experience. By the end of the program, learners will have built and deployed functional AI applications they can adapt for their own projects.

Welcome & Course Orientation Instructor and participant introductions Course goals and objectives Hugging Face at a glance: the “GitHub of AI” Overview of the Hugging Face ecosystem: Hub, Transformers, Datasets, Spaces Software setup check and Colab/Jupyter introduction Hugging Face Hub Tour Navigating the Hub: search, filters, and model cards Popular tasks: NLP, vision, audio, multimodal Understanding model tags, licenses, and intended use Cloning repositories and downloading models locally Getting Started with Pipelines What is a pipeline? Running a sentiment analysis model in 5 lines of code Other built-in pipelines: summarization, translation, question answering Parameters and customization options Performance considerations Practical Text-Based Use Cases Zero-shot classification Summarization and translation Question answering over custom text Hands-on: building a small “document query” notebook Beyond Text: Vision & Audio Models Image classification with ViT Image generation with diffusers (Stable Diffusion Lite) Speech-to-text with Whisper models Hands-on: choose and run one vision or audio task Customizing Pre-Trained Models Changing model configurations Tokenizer tweaks and preprocessing techniques Using pipelines vs. model/tokenizer API directly Exporting and reusing code for automation Building & Sharing a Demo with Hugging Face Spaces Spaces overview: Gradio and Streamlit Creating a simple interface for a pre-trained model Uploading to Spaces for public or private sharing Fine-Tuning Fundamentals Why fine-tune? Benefits over training from scratch Parameter-efficient fine-tuning (LoRA, QLoRA) Overview of the Trainer API Using the `peft` library for LoRA-based tuning Preparing Your Dataset Using the datasets library Loading public datasets from the Hub Cleaning and tokenizing text Train/test splits and evaluation metrics Hands-On Fine-Tuning (Text Model) Selecting a small model (e.g., DistilBERT) Setting training arguments in Trainer Running fine-tuning in Colab Monitoring training progress Saving and evaluating the model Publishing to the Hugging Face Hub Creating a model card Uploading model weights and metadata Versioning and setting permissions Deployment Pathways Inference API basics Using Spaces for deployment Integrating models into Python or JavaScript applications Example: deploy fine-tuned sentiment classifier to Spaces Optimization & Best Practices Reducing model size for faster inference Using quantization and pruning techniques Keeping models updated Managing costs in production environments