# What Hugging Face Actually Is — Hugging Face Transformers

Source: https://www.skillbyai.com/en/hugging-face/hf-ecosystem

> Name the parts of the Hugging Face ecosystem and say which library or service solves which problem.

## A hub plus a set of libraries

Students often say "Hugging Face" as if it were one model. It is really two things. The **Hub** (huggingface.co) is a hosting platform where people publish **models**, **datasets** and **Spaces** (small demo apps), each stored in a Git-based repository with a README called a *model card* or *dataset card*. Around the Hub sit open-source Python libraries. **transformers** gives one consistent API to download, run and fine-tune pretrained models for text, vision and audio. **tokenizers** provides fast tokenizers written in Rust, **datasets** loads and processes data efficiently, **accelerate** runs the same training code on CPU, one GPU or many, and **peft** adds parameter-efficient fine-tuning such as LoRA. The key idea is *pretrained weights plus a standard interface*: instead of training a language model from scratch, which needs huge data and compute, you start from a model someone already trained and adapt it to your task.

## From the Hub to your application

Weights and configs live on the Hub, the transformers library downloads and caches them, and your application calls the loaded model like any Python object.

![Three connected stages: the Hugging Face Hub, the transformers library on your machine, and your application.](assets/figures/hugging-face/section-1-map.svg) — Figure 1.1: Hub -> transformers library -> your application

## Like a public library with a standard card

Think of the Hub as a huge public library where anyone can donate a book, and transformers as a library card that works with every book. You do not need to learn a new rule for each book: you say `from_pretrained("some-id")` and you get a ready-to-use model, just as one card issues any book from the shelf.

## Models are not all "transformers"

The library is named after the Transformer architecture, but the Hub also hosts diffusion models, classic machine-learning models and embedding models used by other libraries such as diffusers and sentence-transformers. Check the model card's *library* tag before assuming `transformers` can load it.

**Quiz:** Which Hugging Face library is mainly responsible for loading and processing datasets efficiently?

- [ ] transformers
- [ ] accelerate
- [x] datasets
- [ ] peft

*Answer:* datasets. `datasets` loads data from the Hub or local files into Arrow-backed tables and provides `map`, `filter` and splits. transformers runs models, accelerate handles devices and peft handles adapter fine-tuning.