In this Specialization, you’ll learn to build and deploy generative AI applications with Hugging Face. You’ll work with models, datasets, and Spaces; run inference with the Pipeline API; prepare text with AutoTokenizer; load models with AutoModel; evaluate model cards; and apply model selection and responsible-use checks.

You’ll then turn models into interactive applications with Gradio. You’ll load and preprocess datasets, fine-tune transformer models with the Trainer API, evaluate results, publish model cards, build interfaces with gr.Interface and gr.Blocks, create streaming multi-turn chatbots with gr.ChatInterface, and deploy applications to Hugging Face Spaces. You’ll also configure hardware and secrets, consider cost and performance, and query deployed apps with the Gradio Python client.

The final course extends these skills to multimodal and agentic AI. You’ll use CLIP and vision-language models for visual question answering, image captioning, and document understanding, then work with Whisper, Diffusers, LoRA, multimodal RAG, smolagents, and MCP. You’ll also apply safety filtering, adversarial testing, and failure-mode documentation for responsible deployment.

Projet d'apprentissage appliqué

In this Specialization you'll work through applied development scenarios across the Hugging Face ecosystem. You’ll choose models from the Hub using model-card evidence, prepare datasets, fine-tune a pretrained transformer with the Trainer API, evaluate results, and publish a model card. You’ll build Gradio interfaces with gr.Interface and gr.Blocks, create streaming multi-turn chatbots with gr.ChatInterface, and deploy applications to Hugging Face Spaces with hardware configuration and secrets.

You’ll also build multimodal workflows with CLIP and vision-language models for visual question answering, image captioning, and document understanding. Additional work covers Whisper audio transcription, Diffusers image generation, decisions around LoRA fine-tuning versus multimodal RAG, agentic workflows with smolagents and MCP, and responsible deployment using ShieldGemma 2, adversarial testing, and failure-mode documentation.