Today's AI development is SLM: Small, fast AI, mastering small language models.
Preface to the book: Why Small Language Models (SLMs) Now?
Part 1: SLM, the Beginning of a New Era
Chapter 1: The Era of Large Language Models (LLMs) and the Emergence of SLMs
1.1.
The LLM That Changed the World: ChatGPT, GPT-4, and Beyond
1.2. The Shadow of LLM: High Costs, Slow Reasoning, and Difficulty in Control
1.3.
The Rise of Alternatives: What is the Small Language Model (SLM)?
1.4. LLM vs. SLM: What, When, and How to Choose?
1.5.
A look at representative SLM models
Chapter 2: How SLM Works: How Does It Differ from LLM?
2.1.
Revisiting the Core Principles of Transformer Architecture
2.2.
The Art of Making Small: Innovations in Model Architecture
2.3.
Efficient Learning: The Importance of High-Quality Data
2.4.
Tokens, Embeddings, and Attention: Understanding the Internals of SLM
Part 2: Hands-on with SLM (Basic Usage)
Chapter 3: My First SLM: Setting Up the Development Environment
3.1.
Required materials: Python, Jupyter Notebook/VS Code
3.2.
Install core libraries: PyTorch, Transformers, and Accelerate
3.3. CUDA Settings for GPU Usage (NVIDIA GPU Users)
3.4.
Get started easily with a cloud/Colab environment.
3.5.
[Hands-on] "Hello, SLM!" - Loading the Model and Generating Your First Text
Chapter 4: Mastering the Hugging Face and SLM Ecosystem
4.1.
Hugging Face: A Playground for AI Developers
4.2.
Model Hub: Find and compare the SLM models you want.
4.3.
Datasets: Exploring Datasets for Fine-Tuning
4.4.
The core of the Transformers library: how to use pipelines
4.5.
[Practice] Creating a Text Summarization and Question-Answering Bot in Just 5 Lines with Pipeline
Part 3: Building Your Own SLM Model (Tuning)
Chapter 5: Fine-Tuning: Injecting My Data into SLM
5.1.
What is fine tuning? Why is it necessary? (General purpose vs. specialized models)
5.2.
Data Preparation: Creating a High-Quality Instruction Dataset
5.3.
Full Fine-tuning: A powerful but heavy method
5.4.
[Practice] Create Your Own Chatbot that Imitates Specific Speech Tone (Full Fine-Tuning)
Chapter 6: Smart Tuning with the Lightweight Tuning Technique (PEFT)
6.1. The Need for Parameter-Efficient Fine-Tuning (PEFT)
6.2.
LoRA (Low-Rank Adaptation): The most popular lightweight tuning technique
6.3.
Easily Understand How LoRA Works
6.4.
[Hands-on] Fine-tuning a low-spec GPU using LoRA
6.5. QLoRA: Tuning with Lower Memory
Chapter 7: Model Quantization: Making It Lighter and Faster
7.1.
What is Quantization? (FP16, INT8, INT4)
7.2.
Why Quantization?: Saving Memory and Speeding Up Inference
7.3.
Quantized loading using the bitsandbytes library
7.4.
[Practice] Loading SLM with 4-bit quantization and comparing resource usage.
Part 4: SLM, Practical Application Development and Deployment
Chapter 8: Synergy Between RAG (Research Augmentation Generation) and SLM
8.1.
Limitations of Language Models: Hallucinations and Lack of Up-to-Date Information
8.2. What is Retrieval-Augmented Generation (RAG)?
8.3.
The Role of Vector DBs and Embedding Models (feat. FAISS, ChromaDB)
8.4.
[Hands-on] Creating a Document-Based Q&A Chatbot by Combining SLM and RAG
Chapter 9: SLM Application Deployment Strategies
9.1.
Optimal Strategy for Each Deployment Scenario
9.2.
Local/On-Premises Deployment: Leveraging Ollama and vLLM
9.3.
Cloud Server Deployment: Building Your Own API Server with FastAPI
9.4.
Edge Device Deployment: Introducing Llama.cpp and the MLX Framework
9.5.
[Practice] Serving a Fine-Tuned Model as an API Using FastAPI
Chapter 10: Responsible Use of SLM and Future Prospects
10.1 AI Ethics: Issues of Bias, Fairness, and Transparency
10.2. SLM and Data Privacy
10.3. The Future of SLM: Smaller, Smarter, and More Multimodal
10.4.
Growing with the open source community
supplement
Appendix A: Glossary of Terms
Appendix B: List of Useful Tools and Libraries
Appendix C: Situation-Specific Recommended SLM Models

Low cost: Operating costs are dozens of times lower than those of an LLM.
High-speed response: Fast inference speed suitable for real-time services.
Enhanced security: Runs on local and intranet networks without data leakage.

Custom optimization: Create AI tailored to specific domains, speech patterns, and business processes.

Developers around the world are already leveraging SLM in a variety of areas, including customer service chatbots, on-device assistants, industrial data analytics, and real-time voice assistants.
An era has begun where we can create “the AI ​​I want, in the environment I want,” without the need for large-scale computational resources.

This book guides you step-by-step through implementing SLM from start to finish, from the basic concepts to the latest practices, fine-tuning, lightweighting, RAG combination, and deployment in various environments.
Beyond simply following code, we've designed it to help you develop practical skills that allow you to design and optimize your own code.