FORMATION IA
Hugging Face 101 : Modèles ouverts pour les ingénieurs
Maîtrisez Hugging Face Hub, déployez des modèles ouverts et choisissez le bon modèle pour vos cas métier.
Ce qu'elle couvre
Cette formation pratique introduit l'écosystème Hugging Face de A à Z pour les ingénieurs. Les participants apprennent à rechercher, évaluer et exécuter des modèles open source avec la bibliothèque Transformers, à déployer des endpoints d'inférence et à publier des démos interactives avec Spaces. À l'issue de la formation, ils sont capables de prendre des décisions éclairées sur le choix de modèles en tenant compte des contraintes métier : latence, coût et confidentialité des données.
À l'issue, vous saurez
- Load and run any model from Hugging Face Hub using Transformers pipelines in under 10 lines of Python
- Fine-tune a pre-trained text or vision model on a custom dataset using the Trainer API or PEFT/LoRA
- Deploy a model as a live REST endpoint using Hugging Face Inference Endpoints and call it from an application
- Build and publish a shareable Gradio demo on Hugging Face Spaces within a single session
- Evaluate and compare open models against business criteria (accuracy, latency, privacy, licensing) to make a justified selection
Sujets abordés
- Navigating the Hugging Face Hub: search, filters, model cards, and leaderboards
- Using the Transformers library: pipelines, tokenizers, and model loading
- Fine-tuning pre-trained models with Trainer API and PEFT/LoRA
- Deploying models via Hugging Face Inference Endpoints
- Building and publishing interactive demos with Gradio and Spaces
- Reading and writing model cards for documentation and governance
- Comparing open models vs. proprietary APIs on cost, latency, and privacy
- Quantisation basics: running models efficiently with bitsandbytes and GGUF
Modalité
Typically delivered as a 2-3 day in-person or live-virtual bootcamp with a 70/30 hands-on to instruction ratio. Each session includes guided lab notebooks hosted on Google Colab or a pre-configured cloud environment. Participants receive access to a shared Hugging Face organisation for collaboration. A take-home capstone project (deploying a task-specific model end-to-end) is included. Remote delivery works well; in-person adds value for team alignment discussions around model selection.
Ce qui fait que ça marche
- Pairing each concept with a real internal use case so participants immediately see business relevance
- Establishing a shared team Space and model registry during training to build collaborative habits from day one
- Including a model-selection rubric workshop so engineers can justify open-model choices to non-technical stakeholders
- Following up with a 2-week async check-in where participants share their capstone results and blockers
Erreurs fréquentes
- Pulling the largest available model by default without checking inference cost, latency, or licence compatibility
- Skipping model cards and leaderboard context, leading to poor model-task fit in production
- Treating Hugging Face Inference Endpoints as a production-grade scalable solution without understanding cold-start and rate-limit constraints
- Ignoring quantisation options and attempting to run 7B+ parameter models on CPU, causing frustration and abandonment
Quand NE PAS suivre cette formation
Teams that have already standardised on a single proprietary LLM API (e.g., OpenAI GPT-4o) with no plans to self-host or fine-tune — the open-model tooling overhead adds complexity without payoff for them.
Fournisseurs à considérer
Sources
Cette formation fait partie d'un catalogue Data & IA construit pour les leaders sérieux sur l'exécution. Lancez le diagnostic gratuit pour voir quelles formations sont prioritaires pour votre équipe.