Introduction

Introduction

Pretraining a transformer takes huge datasets and a lot of compute. Fortunately, you rarely need to: companies, universities and individuals share pretrained models openly. The Hugging Face ecosystem has become the standard way to find, download and use them:

• the transformers library: one consistent Python interface to thousands of model architectures (BERT, GPT-2, T5, Llama, ViT, Whisper, CLIP…), built on PyTorch;
• the Model Hub (huggingface.co): hundreds of thousands of pretrained and fine-tuned models, with documentation ("model cards");
• companion libraries: datasets (loading and processing datasets), tokenizers (fast tokenization), peft (LoRA and friends, Training tab topic 5.3), accelerate (multi-GPU training) and evaluate.

Think of the Hub as an enormous app store for AI models, and the transformers library as the universal phone that can run any app from it.