LLMs & GenAI — Roadmap

This is where classical deep learning turns into ChatGPT-class systems. Assumes you've covered Deep Learning first.

1. Foundation Model & Transformer Internals

• The full pipeline: text -> tokenizer -> embeddings -> positional encoding -> transformer blocks -> LM head -> logits -> sampling -> token
• Tokenization: BPE, WordPiece, SentencePiece
• Embeddings and the embedding table
• Positional encoding: RoPE and ALiBi
• KV cache — why it makes autoregressive decoding fast
• Attention variants: MHA, GQA, MQA, MLA
• Sparse and sliding-window attention
• Encoder-only vs decoder-only vs encoder-decoder architectures
• The MLP block: SwiGLU and gating
• Mixture of Experts (MoE) — sparse vs dense models
• Sampling: greedy, temperature, top-k, top-p/nucleus, repetition penalty
• Context window limits and why they exist

2. Training Pipeline

• Pretraining: next-token prediction at scale
• The data pipeline: crawl -> clean -> dedupe -> filter -> tokenize -> pack
• Scaling laws and the Chinchilla compute-optimal result
• Supervised Fine-Tuning (SFT)
• RLHF — Reinforcement Learning from Human Feedback
• Direct Preference Optimization (DPO), IPO, KTO, ORPO, RLOO
• Group Relative Policy Optimization (GRPO)
• Parameter-efficient fine-tuning: LoRA, QLoRA, DoRA, adapters, IA3, prefix/prompt tuning
• Knowledge distillation
• Quantization (int8, int4, GPTQ, AWQ) for inference

3. Prompt Engineering

• Zero-shot, one-shot, few-shot prompting
• Chain-of-Thought and self-consistency
• ReAct (reasoning + acting) prompting
• Structured output (JSON/XML) prompting
• Prompt injection and jailbreaking — attack and defense

4. Retrieval-Augmented Generation (RAG)

• RAG architecture: retriever + generator
• Chunking strategies (fixed, semantic, recursive, parent-child)
• Embedding model selection
• Dense vs sparse retrieval, BM25, hybrid search + Reciprocal Rank Fusion
• Re-ranking: bi-encoders vs cross-encoders
• Query transformation: HyDE, decomposition, step-back, multi-query, RAG-Fusion
• Advanced: Agentic RAG, Corrective RAG (CRAG), Self-RAG, GraphRAG
• Multimodal RAG, SQL RAG, Code RAG
• Evaluating RAG: faithfulness, groundedness, answer relevance, context precision/recall, MRR/NDCG
• Failure modes: hallucination despite correct context, "lost in the middle," stale knowledge

5. Evaluation

• Perplexity and other intrinsic metrics
• LLM-as-judge evaluation
• Benchmark suites (MMLU, HellaSwag, etc.) — what they do and don't tell you
• Evaluating for hallucination, toxicity, bias

6. Multimodal & Generative Models

• Vision-Language Models (VLMs)
• Vision-Language-Action models (VLAs)
• Diffusion models vs autoregressive generation
• Diffusion Language Models (DLMs)

7. Serving & Production (LLMOps)

• Inference optimization: batching, speculative decoding, Paged Attention
• Cost/latency tradeoffs at scale
• Guardrails and safety filtering in production