Large Language Models: Theory and Modern Multimodal Architectures

Transformers and Attention: A Deep Dive

This module builds a rigorous, mechanistic understanding of the transformer architecture — covering self-attention, multi-head attention, positional encoding, and the encoder-decoder structure — grounded in the original 'Attention Is All You Need' paper.

LLM Pretraining: Objectives, Data, and Scale

This module covers how large language models are pretrained from scratch — examining the core training objectives (autoregressive and masked language modeling), how tokenization works, how training data is collected and curated, and how scale in compute, data, and parameters governs model capability.

Modern LLM Architectures: GPT, LLaMA, Mistral and Beyond

This module examines the key architectural innovations that distinguish state-of-the-art LLMs from the original transformer — including grouped query attention, rotary positional embeddings, mixture of experts, and sliding window attention — explaining the motivation and mechanics behind each.

Alignment and Fine-Tuning: RLHF, PEFT, and Instruction Tuning

Pretrained LLMs are powerful but raw — they predict tokens, not answers. This module covers the techniques used to transform base models into helpful assistants: supervised fine-tuning, instruction tuning, RLHF, DPO, and parameter-efficient methods like LoRA.

Vision-Language Models: Architectures and Training

This module examines how vision and language are fused in modern multimodal models — from CLIP's contrastive pretraining to the cross-attention mechanisms in Flamingo, the projection-layer approach in LLaVA, and the unified reasoning of GPT-4V.