NVIDIA GenAI LLM study plan
NVIDIA GenAI LLM · Give each week a clear focus.
8 weeks · About 7 hours per week
Core Machine Learning and AI Knowledge
Derive scaled dot-product attention and multi-head attention on paper, including the softmax and 1/sqrt(d_k) scaling
Implement a multi-head self-attention module from scratch in PyTorch and verify output shapes against a hand-computed 3-token example
Explain and implement causal masking, positional encoding, and a minimal decoder block end to end
Compute parameter counts and KV-cache memory for a hypothetical 7B model and reconcile the estimate against a real config.json
Your checkpoint
A runnable notebook containing a from-scratch transformer decoder block that passes 8 self-written shape/value assertions and a one-page memory-and-parameter budget table for three model sizes.
Core Machine Learning and AI Knowledge & Software Development
Contrast pretraining, SFT, RLHF, and DPO objectives and articulate when each is appropriate
Implement a LoRA adapter layer manually (A/B low-rank matrices, scaling) and merge it back into base weights
Fine-tune a small open model with LoRA/QLoRA on a toy instruction dataset and log loss, learning rate, and VRAM
Diagnose overfitting and catastrophic forgetting by comparing base-model outputs before and after fine-tuning
Your checkpoint
A fine-tuning repo whose README reports base-vs-tuned loss curves, peak VRAM, trainable-parameter percentage, and five side-by-side prompt outputs demonstrating a measurable behavior change.
Software Development & Core Machine Learning and AI Knowledge
Serve a model behind an async FastAPI endpoint with streaming responses, request validation, and timeout handling
Containerize the service with a slim CUDA base image and pin dependencies for reproducibility
Write pytest unit tests plus a mocked-LLM integration test that runs in CI without a GPU
Profile p50/p95 latency and tokens-per-second, then identify the single largest bottleneck
Your checkpoint
A containerized inference service with 10+ passing tests, a CI config that runs them on every commit, and a benchmark report showing p50/p95 latency before and after one optimization.
Experimentation
Formulate a testable hypothesis with a primary metric, guardrail metrics, and a pre-registered analysis plan
Compute required sample size from baseline rate, minimum detectable effect, alpha, and power using both formula and simulation
Design an A/B test for an LLM feature, including unit of randomization and novelty/primacy controls
Identify three ways peeking or optional stopping would invalidate the result and state the mitigation for each
Your checkpoint
A pre-registration document for one A/B test containing the hypothesis, primary and guardrail metrics, sample-size calculation with code, and a stated decision rule.
Experimentation & Data Analysis
Analyze a simulated experiment dataset with a two-proportion z-test, confidence intervals, and effect sizes
Compare frequentist and bootstrap approaches and reconcile any disagreement in the conclusion
Build an LLM-as-judge evaluation harness with a rubric and measure its agreement against human labels using Cohen's kappa
Detect and explain a confound or Simpson's-paradox pattern deliberately planted in the data
Your checkpoint
An experiment readout notebook producing a ship/no-ship recommendation backed by an effect size, confidence interval, and judge-human agreement score above 0.6 kappa.
Data Analysis
Profile a raw text corpus with pandas: token-count distributions, duplicate rate, language mix, and outliers
Implement near-duplicate detection using MinHash or shingled Jaccard similarity and quantify removal impact
Scan for and redact PII patterns with regex plus a validation pass, reporting false-positive rate on a labeled sample
Assemble a train/eval split that avoids leakage and document the filtering pipeline as reproducible code
Your checkpoint
A reproducible data-curation script plus an EDA report covering 5+ quality metrics, two before/after charts, and a documented leakage audit of the resulting splits.
Trustworthy AI
Red-team a deployed endpoint with a 25-prompt adversarial suite covering prompt injection, jailbreaks, and data exfiltration
Measure hallucination rate on a 50-question grounded QA set and classify failures by root cause
Evaluate model outputs for demographic bias using a paired-prompt methodology and report the disparity
Write a model card and a mitigation plan mapping each identified risk to a guardrail, monitor, or documented acceptance
Your checkpoint
A trust-and-safety report with a scored 25-prompt red-team table, a hallucination rate with confidence interval, a bias disparity metric, and a completed model card.
Core Machine Learning and AI Knowledge & Software Development & Experimentation & Data Analysis & Trustworthy AI
Re-derive the five weakest concepts identified from prior weeks' deliverables without notes
Complete two timed full-length practice sets and log error causes by domain
Rebuild one end-to-end mini-project (data curation, fine-tune, serve, evaluate, document risks) from a blank repo
Close every remaining gap flagged as needs-review in the personal error log
Your checkpoint
An end-to-end capstone repo with tests, an evaluation report, and a model card, plus an error log showing every domain above 75% on the final practice set.
Milestones
Week 2: Explain transformer attention, KV-cache cost, and the pretraining/SFT/DPO distinction from memory, and merge a trained LoRA adapter into base weights successfully.
Week 4: Pass a self-graded quiz on statistics fundamentals at 80%+ and defend a pre-registered A/B design against three written objections without notes.
Week 6: Full experiment readout and data-curation report both accepted: recommendation is decision-ready and the leakage audit found zero unintended train/eval overlap.
Week 8: Capstone runs end to end in a clean environment; all five domains score above 75% on timed practice, with no open needs-review items.