Hugging Face LLM Course
by Hugging Face
The maintainers' own thirteen-chapter path from transformer internals to GRPO reasoning models.
Overview
Formerly the Hugging Face NLP Course, this is a thirteen-chapter path, Chapter 0 setup plus Chapters 1 through 12, that teaches modern LLM engineering through Hugging Face's own libraries. Chapters 1 to 4 cover transformer architectures, what actually happens behind the pipeline abstraction, models, tokenizers, handling multiple sequences, optimized inference deployment, and publishing to the Hub with model cards. Chapters 5 to 8 go deeper: the Datasets library, including corpora too big for memory and semantic search with FAISS; the Tokenizers library down to Byte-Pair Encoding, WordPiece and Unigram, building a tokenizer block by block; classical NLP tasks such as token classification, masked language modeling, translation, summarization, question answering and training a causal language model from scratch; and a chapter on debugging the training pipeline and writing a good issue. Chapter 9 builds and shares demos with Gradio. Chapters 10 to 12 are the newest material: curating datasets with Argilla, then chat templates, SFTTrainer, LoRA and evaluation, then building reasoning models through reinforcement learning on LLMs, the aha moment from the DeepSeek R1 paper, GRPO as described in DeepSeekMath, and implementing GRPO in TRL. Each chapter is designed for roughly six to eight hours of work. Hugging Face develops it in the open at github.com/huggingface/course, which carries over 4,100 stars and 1,380 forks and was last pushed in July 2026, so the code tracks current library versions instead of rotting.
At a Glance
- Topic
- Models
- Level
- Intermediate
- Format
- Course
- Cost
- Free
- Duration
- 13 chapters, ~6-8 hours per chapter, self-paced
- Provider
- Hugging Face
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Explain transformer architectures and what happens inside the pipeline abstraction
- ✓Fine-tune pretrained models with the Trainer API and a raw training loop
- ✓Implement Byte-Pair Encoding, WordPiece and Unigram tokenization from the ground up
- ✓Handle datasets too large for memory using streaming and FAISS semantic search
- ✓Build token classification, summarization, translation and question-answering pipelines
- ✓Debug a broken training pipeline and interpret the errors it throws
- ✓Apply LoRA and SFTTrainer to fine-tune an LLM on your own data
- ✓Train reasoning models with GRPO in TRL, following the DeepSeek R1 recipe
- ✓Ship an interactive model demo with Gradio Blocks on the Hub
Highlights
- •Written by the maintainers of Transformers, Datasets, Tokenizers and TRL, so the code matches the libraries
- •Developed in the open: over 4,100 stars and 1,380 forks on GitHub, last pushed July 2026
- •Chapters 11 and 12 cover LoRA, SFTTrainer and GRPO reasoning-model training, material most 2023-era courses never reached
- •Chapter 8 teaches debugging the training pipeline, a topic almost every other course skips
- •Entirely free with no advertising and no paid tier
Who It’s For
Best For
- ✓Python developers moving from calling LLM APIs to training and fine-tuning them
- ✓ML engineers who need the Hugging Face stack as a working tool, not a demo
- ✓Practitioners who want to understand tokenization and attention below the abstraction
- ✓Anyone preparing to fine-tune an open-weight model with LoRA on their own data
Prerequisites
- •Strong working knowledge of Python
- •An introductory deep learning course such as fast.ai's Practical Deep Learning for Coders
- •PyTorch or TensorFlow experience is helpful but explicitly not required
FAQ
What is Hugging Face LLM Course?
Hugging Face's free, open-source course covering transformers end to end: architecture, tokenizers, fine-tuning, datasets, Gradio demos, and in the newest chapters LoRA, supervised fine-tuning with TRL, and training reasoning models with GRPO. Written and maintained by the people who ship the libraries, and updated in the open on GitHub rather than frozen at publication.
Is Hugging Face LLM Course free?
Hugging Face LLM Course is free to access.
What level is Hugging Face LLM Course for?
Hugging Face LLM Course is aimed at a intermediate audience. Recommended background: Strong working knowledge of Python, An introductory deep learning course such as fast.ai's Practical Deep Learning for Coders, PyTorch or TensorFlow experience is helpful but explicitly not required.
How long does Hugging Face LLM Course take?
Expect roughly 13 chapters, ~6-8 hours per chapter, self-paced. Most learners work through it at their own pace.
What will I learn from Hugging Face LLM Course?
You'll learn: Explain transformer architectures and what happens inside the pipeline abstraction; Fine-tune pretrained models with the Trainer API and a raw training loop; Implement Byte-Pair Encoding, WordPiece and Unigram tokenization from the ground up; Handle datasets too large for memory using streaming and FAISS semantic search; Build token classification, summarization, translation and question-answering pipelines; Debug a broken training pipeline and interpret the errors it throws; Apply LoRA and SFTTrainer to fine-tune an LLM on your own data; Train reasoning models with GRPO in TRL, following the DeepSeek R1 recipe; Ship an interactive model demo with Gradio Blocks on the Hub.
Topics
Sources
This page was written from 4 sources, 3 on domains other than huggingface.co.