NCA-GENL Study Plan¶

6-Week Comprehensive Study Schedule¶

Week 1: LLM Fundamentals - Architecture and Theory¶

Day 1-2: Transformer Architecture¶

Study the original transformer architecture (encoder, decoder, encoder-decoder)
Understand self-attention mechanism (Q, K, V matrices)
Learn multi-head attention and why it matters
Study positional encoding methods (absolute, relative, rotary - RoPE)
Understand layer normalization and residual connections
Review feed-forward network role in transformer blocks
Hands-on: Explore transformer visualization tools
Review Notes: notes/01-transformer-architecture.md

Day 3-4: Tokenization and Embeddings¶

Study Byte-Pair Encoding (BPE) algorithm
Learn WordPiece tokenization (used by BERT)
Understand SentencePiece (used by LLaMA, T5)
Learn how vocabulary size affects model performance
Study embedding layers and how tokens become vectors
Understand special tokens (CLS, SEP, BOS, EOS, PAD)
Hands-on: Experiment with different tokenizers on Hugging Face
Review Notes: notes/01-transformer-architecture.md

Day 5-6: Model Families and Pre-training¶

Study GPT family (decoder-only, autoregressive generation)
Learn BERT family (encoder-only, bidirectional understanding)
Understand T5 family (encoder-decoder, text-to-text)
Study LLaMA, Mistral, and Mixtral architectures
Learn NVIDIA Nemotron model family
Understand pre-training objectives (CLM, MLM, span corruption)
Study scaling laws and emergent abilities
Hands-on: Load and compare different model architectures on Hugging Face
Review Notes: notes/01-transformer-architecture.md

Day 7: Week 1 Review¶

Create summary of key architecture concepts
Quiz yourself on attention mechanism details
Verify understanding of model type differences (encoder vs decoder)
Review any weak areas identified during the week

Week 2: Prompt Engineering¶

Day 8-9: Core Prompting Techniques¶

Study zero-shot prompting and when it works best
Learn one-shot and few-shot prompting strategies
Understand in-context learning capabilities and limitations
Practice writing effective task instructions
Study prompt template design patterns
Hands-on: Test zero-shot vs few-shot on NVIDIA AI Playground
Review Notes: notes/02-prompt-engineering.md

Day 10-11: Advanced Prompting and Output Control¶

Study chain-of-thought (CoT) prompting technique
Learn self-consistency (multiple CoT paths)
Understand system prompts and role-based instructions
Study temperature, top-k, and top-p sampling parameters
Learn output formatting (JSON, structured responses)
Practice designing prompts for different task types
Hands-on: Experiment with sampling parameters using NIM APIs
Review Notes: notes/02-prompt-engineering.md

Day 12-13: Prompt Security and Best Practices¶

Study prompt injection attacks (direct and indirect)
Learn prompt injection mitigation strategies
Understand prompt-response evaluation methods
Study task decomposition for complex problems
Learn about stop sequences and response truncation
Hands-on: Test prompt injection scenarios and defenses
Review Notes: notes/02-prompt-engineering.md

Day 14: Week 2 Review¶

Practice writing prompts for 5 different task types
Test understanding of sampling parameter effects
Review prompt security concepts
Identify knowledge gaps for further study

Week 3: RAG and Vector Databases¶

Day 15-16: RAG Architecture Fundamentals¶

Study the complete RAG pipeline (ingest, embed, index, retrieve, generate)
Understand why RAG reduces hallucination
Learn embedding models and how they create vector representations
Study similarity metrics (cosine similarity, dot product, Euclidean)
Understand bi-encoder vs cross-encoder models
Learn about NVIDIA NeMo Retriever components
Hands-on: Build a simple RAG pipeline with LangChain
Review Notes: notes/03-rag-vector-databases.md

Day 17-18: Vector Databases Deep Dive¶

Study FAISS - index types (Flat, IVF, HNSW), GPU acceleration
Learn Milvus - features, scaling, production deployment
Understand other options (Pinecone, Weaviate, Chroma, Qdrant)
Study indexing methods (HNSW, IVF, PQ) and their trade-offs
Learn about metadata filtering in vector search
Understand approximate nearest neighbor (ANN) search
Hands-on: Set up FAISS and Milvus, compare search performance
Review Notes: notes/03-rag-vector-databases.md

Day 19-20: Chunking, Indexing, and Retrieval Quality¶

Study chunking methods (fixed-size, recursive, semantic, document-aware)
Understand chunk size trade-offs (precision vs context)
Learn about chunk overlap and why it matters
Study hybrid search (dense + sparse retrieval)
Understand re-ranking and its impact on result quality
Learn context window management strategies
Hands-on: Experiment with different chunking strategies
Review Notes: notes/03-rag-vector-databases.md

Day 21: Week 3 Review¶

Design a complete RAG pipeline for a given use case
Compare vector database options for different scenarios
Practice chunking strategy selection
Review retrieval quality evaluation methods

Week 4: Fine-Tuning and Customization¶

Day 22-23: Fine-Tuning Fundamentals¶

Study full fine-tuning process and requirements
Understand GPU memory requirements for fine-tuning
Learn about catastrophic forgetting and how to prevent it
Study training data preparation and quality requirements
Understand evaluation metrics (perplexity, BLEU, ROUGE)
Learn about NVIDIA NeMo Framework for training
Hands-on: Review NeMo Framework documentation and tutorials
Review Notes: notes/04-fine-tuning-customization.md

Day 24-25: PEFT Methods (LoRA, QLoRA, Adapters)¶

Study LoRA - low-rank matrices, rank selection, target modules
Learn QLoRA - 4-bit quantization + LoRA, NF4 format
Understand other PEFT methods (prefix tuning, adapters, IA3)
Compare PEFT methods by memory efficiency, quality, and complexity
Study when to choose each PEFT method
Learn about adapter merging and multi-adapter serving
Hands-on: Fine-tune a small model with LoRA using NeMo or Hugging Face
Review Notes: notes/04-fine-tuning-customization.md

Day 26-27: Alignment Methods (RLHF, DPO)¶

Study supervised fine-tuning (SFT) as alignment baseline
Learn RLHF pipeline (reward model, PPO optimization)
Understand DPO as simplified alternative to RLHF
Study preference data collection and formatting
Learn about KL divergence penalty in alignment
Understand when to use SFT vs RLHF vs DPO
Hands-on: Review NeMo alignment documentation
Review Notes: notes/04-fine-tuning-customization.md

Day 28: Week 4 Review¶

Compare all fine-tuning approaches (full, LoRA, QLoRA, RLHF, DPO)
Practice selecting fine-tuning methods for different scenarios
Create decision tree for fine-tuning method selection
Review GPU memory requirements for different approaches

Week 5: Deployment, Inference, and Ethics¶

Day 29-30: NVIDIA Deployment Stack¶

Study NVIDIA NIM - architecture, deployment, API
Learn TensorRT-LLM - optimization techniques, quantization
Understand Triton Inference Server - multi-framework serving
Study the relationship between NIM, TensorRT-LLM, and Triton
Learn about containerized model deployment
Hands-on: Deploy a model using NVIDIA NIM
Review Notes: notes/05-deployment-inference.md

Day 31-32: Inference Optimization¶

Study quantization methods (FP8, INT8, INT4, AWQ, GPTQ)
Learn KV-cache and paged attention concepts
Understand batching strategies (static, dynamic, continuous)
Study model parallelism (tensor, pipeline)
Learn speculative decoding for faster generation
Understand throughput vs latency trade-offs
Hands-on: Compare inference performance with different quantization
Review Notes: notes/05-deployment-inference.md

Day 33-34: Ethics and Responsible AI¶

Study bias sources in LLMs (data, training, output)
Learn NVIDIA NeMo Guardrails (topical, safety, fact-checking rails)
Understand hallucination types and mitigation strategies
Study privacy considerations (PII handling, data governance)
Learn about red-teaming and adversarial testing
Understand regulatory frameworks (EU AI Act awareness)
Hands-on: Configure NeMo Guardrails for a chatbot
Review Notes: notes/06-ethics-responsible-ai.md

Day 35: Week 5 Review¶

Create comparison chart of NVIDIA deployment tools
Review quantization methods and trade-offs
Summarize responsible AI practices
Identify remaining knowledge gaps

Week 6: Review, Practice, and Exam Preparation¶

Day 36-37: Comprehensive Review¶

Review all notes files and fact sheet
Create flashcards for key concepts and NVIDIA tools
Review quick reference tables (architecture, PEFT, quantization)
Study NVIDIA tool selection guide
Revisit any topics with low confidence

Day 38-39: Scenario Practice¶

Work through all scenarios in scenarios.md
Practice reasoning through trade-off questions
Design RAG pipelines for different use cases
Select fine-tuning methods for various requirements
Practice NVIDIA tool selection for deployment scenarios

Day 40-41: Mock Assessment and Gap Analysis¶

Take any available practice assessments
Time yourself (60 minutes for 50-60 questions)
Review all incorrect answers thoroughly
Focus additional study on weak areas
Re-read NVIDIA documentation for missed topics

Day 42: Final Preparation¶

Quick review of fact sheet and key tables
Review common exam pitfalls from strategy guide
Verify understanding of NVIDIA ecosystem tools
Light review only - avoid cramming new material
Prepare exam environment (online proctoring setup)
Get a good night's rest before exam day

Supplementary Study Activities¶

Weekly Labs (Pick at Least One Per Week)¶

Lab 1: Explore model architectures on Hugging Face Model Hub
Lab 2: Build a chatbot with system prompts and sampling controls
Lab 3: Create a RAG pipeline with FAISS and LangChain
Lab 4: Fine-tune a model with LoRA on a custom dataset
Lab 5: Deploy a model with NVIDIA NIM
Lab 6: Implement NeMo Guardrails for safety

Ongoing Activities¶

Read NVIDIA developer blog posts weekly
Follow GTC presentations on generative AI topics
Practice explaining concepts to solidify understanding
Join NVIDIA developer forums for community support