Large Language Models: Deep Dive

From embeddings and attention to the full transformer block. The architecture behind every modern LLM.

• 3Blue1Brown: Neural Networks + Transformers (YouTube)videofree
• The Illustrated Transformer (Jay Alammar)articlefree
• Attention Is All You Need (paper)articlefree

• Explain self-attention with a worked example
• Diagram a transformer block
• Implement scaled dot-product attention in NumPy

Follow Karpathy's nanoGPT to implement, train, and sample from a small GPT. The single best way to truly understand LLMs.

• Karpathy: Let's build GPT (YouTube)videofree
• nanoGPT (GitHub)repofree
• Karpathy: Neural Networks Zero to Herocoursefree

• Train a character-level GPT
• Implement multi-head attention
• Sample coherent text from your model

The full training pipeline: next-token pretraining, supervised fine-tuning, and alignment with RLHF/DPO.

• HuggingFace LLM Course (free)coursefree
• Karpathy: State of GPT (talk)videofree
• InstructGPT / RLHF (paper)articlefree

• Explain pretraining vs. SFT vs. RLHF
• Describe reward modeling
• Compare RLHF and DPO

Adapt open models on a single GPU with LoRA/QLoRA, dataset prep, and evaluation of the result.

• HuggingFace PEFT Docsdocfree
• QLoRA (paper)articlefree
• HuggingFace: Fine-tune an LLM (recipe)articlefree

• Prepare an instruction dataset
• Fine-tune a 7B model with QLoRA
• Evaluate before/after on held-out data

Benchmarks, LLM-as-judge, bias, toxicity, and building task-specific eval harnesses.

• HuggingFace Evaluate Docsdocfree
• Open LLM Leaderboardlinkfree
• Holistic Evaluation of Language Models (HELM)articlefree

• Build a task-specific eval set
• Use LLM-as-judge with rubrics
• Document a bias/safety finding

Quantization, vLLM, KV-cache, batching, and cost/latency tradeoffs for real deployments.

• vLLM Docsdocfree
• HuggingFace Text Generation Inferencedocfree
• Ollama (run models locally)repofree

• Serve a model with vLLM
• Quantize and measure latency change
• Estimate cost per 1M tokens