Open-Source AI Cookbook

The Cookbook is organized as a browsable index of individual notebooks rather than a linear course, which makes it a reference you return to rather than something you finish. The fine-tuning and post-training cluster is the deepest: supervised fine-tuning of code LLMs on a single GPU, GRPO with TRL, online GRPO training with vLLM, advanced GRPO reward design, DPO and MPO for vision-language models, function-calling fine-tuning on the xLAM dataset, PEFT prompt tuning, multi-config SFT with RapidFire AI, and vision fine-tuning across SmolVLM, Granite Vision, ViT and DETR. The retrieval cluster covers advanced RAG, agentic RAG, multi-agent RAG, multimodal RAG with document retrieval plus rerankers and VLMs, RAG evaluation, semantic reranking and caching, and integrations with Milvus, Elasticsearch, MongoDB, Neo4j knowledge graphs, LlamaIndex and LangChain. The agents cluster includes text-to-SQL agents, data-analyst agents, multi-agent web assistants and smolagents micro-agent architectures. Evaluation and reliability get real coverage too — LLM-as-judge, judging AI search engines, dataset quality issues, active-learning annotation, preference-dataset generation with distilabel, structured generation, PII detection at an LLM gateway, and Phoenix observability on Spaces. Everything runs on open models and open tooling (Transformers, TRL, PEFT, vLLM, Gradio, Argilla), which makes the notebooks adaptable to self-hosted stacks. It is Apache-2.0, maintained in the huggingface/cookbook repository with over 1,300 commits, and explicitly community-driven with contributions accepted.

Topic

Fine-Tuning

Level

Intermediate

Format

Tutorial

Cost

Free

Duration

~45-90 minutes per notebook; 70+ notebooks, entirely self-paced

Provider

Hugging Face

Hands-on

Yes — code/exercises

Certificate

None

• ✓Run GRPO reinforcement post-training with TRL, including online training served by vLLM
• ✓Design advanced reward functions for reasoning-focused GRPO runs rather than copying a default
• ✓Fine-tune vision-language models with SFT, DPO and MPO on SmolVLM and Granite Vision
• ✓Fine-tune a code LLM on a single GPU and tune function calling on the xLAM dataset
• ✓Build multimodal RAG that combines document retrieval, a reranker and a vision-language model
• ✓Evaluate RAG systems and use LLM-as-judge, including judging AI search engines
• ✓Assemble agentic pipelines — text-to-SQL, data analyst, multi-agent web assistant — with smolagents
• ✓Generate and clean preference datasets with distilabel, and add observability with Phoenix

• •Every recipe is a runnable notebook against open models, so nothing is locked behind one vendor's hosted API.
• •It tracks the ecosystem rather than lagging it — current entries include DSPy GEPA optimization, online GRPO with vLLM in TRL, and multi-config SFT with RapidFire AI.
• •Unusually strong on the unglamorous half of the job: dataset quality issues, active-learning annotation, preference-data generation, PII detection at an LLM gateway.
• •Vision-language post-training is covered properly (SFT, DPO, MPO, GRPO across four model families), which most RAG-and-agents tutorials skip entirely.
• •Written by the maintainers of the libraries it uses — Transformers, TRL, PEFT — plus community contributors, so the code matches current library APIs.
• •Apache-2.0 and openly contributed, with 1,300+ commits in huggingface/cookbook; you can fork a recipe as a starting point without licensing friction.

This page was written from 3 sources, 2 on domains other than huggingface.co.

1. 1.huggingface.co — indexvendor
2. 2.github.com — cookbook
3. 3.github.com — en