Hugging Face cheat sheet

1. 1The HubThe shared home for models, datasets and demo apps, and the account and tooling you need to use it from code.7 concepts

   1. 1Hub repositoriesEvery model, dataset and Space on the Hub is a Git-based repository with an id of the form owner/name, such as Qwen/Qwen3-8B.
   2. 2The hf CLI and loginThe hf command line tool (the successor to huggingface-cli) logs in, downloads and uploads files.
   3. 3Model cardsThe README of a model repo: what the model is for, how it was trained, its evaluation results, limitations and intended use, plus YAML metadata such as licence…
   4. 4Licences and gated modelsEach repo declares a licence, such as Apache 2.0, MIT or a custom model licence with usage conditions.
   5. 5SpacesSpaces host small apps, usually built with Gradio or as a Docker container, directly from a Hub repo, on free CPU or paid GPU hardware.
   6. 6The local cacheDownloads are stored once in a shared cache (by default under ~/.cache/huggingface) and reused by every library.
   7. 7huggingface_hub in PythonThe huggingface_hub library downloads single files or whole repos, creates repos, uploads folders and searches the Hub from Python.

   7 in this chapterOpen this chapter
2. 2Running models with transformersThe transformers library: from a one-line pipeline to loading the tokeniser and model yourself and generating text.7 concepts

   1. 1pipelinepipeline(task, model=...) bundles preprocessing, the model and postprocessing for a task such as text generation, classification, summarisation or speech…
   2. 2Auto classesAuto classes read a repo's config and build the right architecture.
   3. 3dtype and device_mapdtype="auto" loads weights in the precision they were saved in, often bfloat16, instead of float32.
   4. 4generatemodel.generate produces tokens one at a time.
   5. 5Chat templatesEach chat model expects messages formatted with its own special tokens.
   6. 6Model outputs and headsCalling a model returns an output object: logits for models with a task head, last_hidden_state for a bare AutoModel, and loss when you pass labels.
   7. 7safetensorsThe default weight format on the Hub: a simple file of tensors with a JSON header, which loads fast and cannot run code.

   7 in this chapterOpen this chapter
3. TokenisersHow text becomes the integer ids a model reads, and the settings that matter when you batch it.6 conceptsOpens with credits
4. DatasetsThe datasets library loads data from the Hub or local files into fast, memory-mapped tables and transforms it in batches.7 conceptsOpens with credits
5. Fine-tuning with TrainerTraining a pretrained model further on your own labelled data with the high-level Trainer API.6 conceptsOpens with credits
6. LoRA, quantisation and accelerateFitting fine-tuning and inference of large models onto modest hardware, and scaling the same code up.6 conceptsOpens with credits
7. Evaluation and embeddingsMeasuring whether a model is good enough, and turning text into vectors for search and similarity.6 conceptsOpens with credits
8. Inference and sharingRunning models without managing GPUs, serving them yourself, and publishing what you trained.6 conceptsOpens with credits
9. 9Glossary, A to ZEvery concept in the sheet and the smaller terms it uses, in one list.75 terms

The Hub, transformers, tokenisers, datasets, fine-tuning, LoRA, quantisation, inference and embeddings with the Hugging Face libraries.