Large Language Models (LLM) Course: From Fundamentals to Production

Topics covered

• •Refresh the mathematical minimum that deep learning is built on.
• •Make sure you're comfortable with the Python ML stack: NumPy, Pandas, Scikit-learn, PyTorch.
• •Understand how neural networks are built and trained: layers, backpropagation, regularization.

Topics covered

• •Understand the evolution from encoder-decoder transformers to decoder-only models (GPT).
• •Understand how tokenization turns text into numbers and how it affects quality.
• •Master the attention mechanism (self-attention) and its role in handling long-range dependencies.

Topics covered

• •Understand how LLM pretraining works: data scale, parallelism, optimization, monitoring.
• •Master pretraining data preparation: curation, cleaning, deduplication, quality filtering.
• •Understand the structure of post-training datasets: storage formats and chat templates.

Topics covered

• •Master supervised fine-tuning (SFT): full fine-tuning vs. LoRA/QLoRA, key parameters.
• •Understand preference alignment: DPO, reward models, RL methods (GRPO, PPO).
• •Learn to scale training (DeepSpeed/FSDP) and monitor it.

Topics covered

• •Understand the principles of quantization: precision levels, absmax, and zero-point.
• •Master practical formats and methods: GGUF/llama.cpp, GPTQ/EXL2, AWQ, SmoothQuant/ZeroQuant.
• •Study inference optimizations: Flash Attention, KV-cache (MQA/GQA), speculative decoding.

Topics covered

• •Learn to run LLMs: API vs. local deployment, prompt engineering, structured output.
• •Build a vector store and a full RAG pipeline, and know how to evaluate it.
• •Master advanced RAG: query construction, tools, re-ranking, DSPy.

Topics covered

• •Understand the LLM-specific vulnerabilities that set them apart from ordinary software.
• •Master the types of prompt attacks: injection, data/prompt leaking, jailbreaks.
• •Understand attacks on training: data poisoning and backdoors.