dive-into-llms

Dive into LLMs is a series of programming practice tutorials expanded from the lecture notes of Shanghai Jiao Tong University's courses "Frontier Technologies in Natural Language Processing" (NIS8021) and "Artificial Intelligence Security Technologies" (NIS3353). Developed by Professor Zhang Zhuosheng's team, it is a public-welfare, completely free project aimed at helping learners quickly get started with large language models through simple hands-on practice, better enabling course design or academic research. The tutorial consists of 11 chapters, each providing slides (PDF), tutorial documentation (README), and runnable Jupyter Notebook scripts: 1. Fine-tuning and Deployment: A guide to fine-tuning and deploying pretrained models, covering model selection, task-specific fine-tuning, and demo deployment. 2. Prompt Learning and Chain-of-Thought: A guide to LLM API calls and inference, exploring prompt engineering and chain-of-thought techniques. 3. Knowledge Editing: Methods and tools for editing language models, enabling the editing and verification of specific knowledge. 4. Mathematical Reasoning: Teaching LLMs mathematical reasoning capabilities by distilling a mini R1 model. 5. Model Watermarking: Techniques for embedding human-invisible watermarks into content generated by language models. 6. Jailbreak Attacks: Understanding how jailbreak attacks breach LLM security defenses, approaching security from an attack perspective. 7. LLM Steganography: Enabling LLMs to carry hidden information recognizable only by specific recipients while responding fluently. 8. Multimodal Models: Exploring how multimodal large language models achieve more powerful multimodal understanding and generation capabilities. 9. GUI Agents: Enabling AI agents to perform GUI operation tasks such as ordering food, replying to messages, and shopping price comparison on behalf of users. 10. Agent Safety: Discussing risk threats and security awareness of LLMs in open agent scenarios. 11. RLHF Safety Alignment: An experimental guide to RLHF based on PPO, covering safety alignment practices in reinforcement learning from human feedback. Additionally, the project collaborates with the Huawei Ascend community to offer a localized public-welfare tutorial titled "Full-Process LLM Development," built on Ascend foundational hardware and software. It includes PPTs, lab manuals, and videos, divided into beginner, intermediate, and advanced series, catering to diverse LLM practice needs and providing a progressive full-process development guide from quick start, to applying Ascend-supported models, to migrating and tuning new models. The project is jointly contributed by multiple teams from Shanghai Jiao Tong University, the National University of Singapore, and the Huawei Ascend community. Contributions and improvements via Issues and Pull Requests are welcome.