LLMs From Scratch: Build a GPT-like Model with Your Own Hands

Topics covered

• •Understand what an LLM is and how it differs from classic NLP systems.
• •Break down the stages of a model's lifecycle: pretraining → fine-tuning.
• •Get an overview of the course roadmap: which components we'll build and in what order.

Topics covered

• •Implement a simple regex-based tokenizer and a token→ID vocabulary.
• •Understand the purpose of special tokens and the principle behind Byte Pair Encoding (BPE).
• •Build a dataset and a sliding-window DataLoader for the 'predict the next token' task.

Topics covered

• •Understand why the transformer needs attention and what the limitations of architectures without it are.
• •Implement, step by step: simple attention without weights → self-attention with trainable Q/K/V matrices → causal attention → multi-head attention.
• •Understand the scaling of dot products and the role of dropout.

Topics covered

• •Assemble the full GPT-2 small model (~124M parameters) from its individual components.
• •Implement LayerNorm, the GELU activation, the feed-forward network, and residual connections.
• •Assemble the transformer block and the full `GPTModel`; count the number of parameters.

Topics covered

• •Implement the standard quality metrics for a generative model: cross-entropy and perplexity.
• •Write a full pretraining loop with a train/val split and periodic sample generation.
• •Learn decoding strategies: temperature scaling and top-k sampling.

Topics covered

• •Understand the difference between classification fine-tuning and instruction tuning.
• •Prepare the SMS spam dataset: class balancing, padding, DataLoader.
• •Replace GPT's output head with a classification head and choose which layers to fine-tune.

Topics covered

• •Understand why supervised instruction fine-tuning (SFT) is needed and how it differs from classification.
• •Prepare an instruction dataset: a prompt template (Alpaca-style), dynamic padding, loss masking.
• •Fine-tune GPT-2 medium (355M) on 1,100 'instruction → response' pairs.