NVIDIA Certified Associate – Generative AI LLMs (NCA-GENL)

Learning plan

About 40 hours of focused study, split into 5 phases. Pick your weekly pace and the timeline adjusts.

Estimated finish: 4 weeks

What the exam covers

Source: official exam guide. Weights can change, so check before you start.

Before you start

• Working Python and basic NumPy/pandas
• A free Google Colab or Kaggle account for GPU notebooks, and a free NVIDIA developer account for build.nvidia.com
• NVIDIA lists one prerequisite: a basic understanding of generative AI and LLMs

1. Weeks 1–2 · 12 h

   ML, deep learning and transformer fundamentals (largest domain)

   • Explain gradient descent, loss functions, overfitting and regularisation, and train a small network end to end
   • Walk through the transformer, covering tokenisation, embeddings, positional encoding, self-attention and encoder vs decoder vs encoder-decoder
   • Compare pre-training, instruction tuning, RLHF and parameter-efficient fine-tuning (LoRA)

   Hands-on: Run the Fundamentals of Deep Learning notebooks on a free Colab GPU through the image-classification and transfer-learning sections, recording validation accuracy before and after augmentation.

   Use: NVIDIA DLI Fundamentals of Deep Learning (open-source materials) · The Illustrated Transformer (Jay Alammar) · Hugging Face LLM Course

2. Weeks 2–3 · 10 h

   Data analysis and experimentation

   • Clean and explore a text dataset (deduplication, length distribution, label balance) with pandas, or with cuDF on GPU
   • Design an experiment varying learning rate, batch size or prompt template, and compare runs fairly
   • Choose evaluation metrics (perplexity, BLEU, ROUGE, F1, human eval) for generation vs classification tasks

   Hands-on: Fine-tune a small model with LoRA on a public classification dataset in Colab. Run three learning rates and report accuracy and F1 for each in one table.

   Use: Hugging Face LLM Course · Hugging Face PEFT docs

3. Weeks 3–4 · 10 h

   Building LLM applications with NVIDIA tooling

   • Apply zero-shot, few-shot and chain-of-thought prompting and control output with temperature and top-p
   • Build a retrieval-augmented generation pipeline (chunking, embeddings, vector search, prompt assembly) with LangChain
   • Call NVIDIA NIM endpoints from build.nvidia.com, and know what NeMo, NIM and TensorRT-LLM each do

   Hands-on: Build a small RAG app over a set of PDFs that uses an embedding NIM and an LLM NIM from build.nvidia.com, then compare answers with and without retrieval on 10 questions.

   Use: NVIDIA DLI Rapid Application Development With LLMs · NVIDIA API catalog (build.nvidia.com) · NVIDIA NIM docs

4. Week 4 · 3 h

   Trustworthy AI

   • Name sources of bias in training data and LLM output, and mitigation steps
   • Explain privacy, consent and transparency concerns, and how NeMo Guardrails restricts topics and outputs

   Use: NeMo Guardrails on GitHub

5. Week 4 · 5 h

   Practice and review

   • Re-quiz yourself on Core ML (30%) and Software Development (24%), which make up over half the exam
   • Time yourself on a 50-question set at about one minute per question

   Use: NCA-GENL exam page and blueprint

You're ready to book when…

• You can sketch the transformer block and explain attention without notes
• You have built a working RAG pipeline and a LoRA fine-tune yourself
• You can name the right metric for summarisation, translation and classification tasks
• You finish timed practice sets of 50 questions in under 60 minutes, scoring 80% or higher

Exam tips

• The exam is 50-60 multiple-choice questions in 60 minutes, so keep moving at about a minute per question.
• The certification is valid for two years, and you recertify by retaking the exam.