Objective Coverage

Every published NCP-GENL exam objective, matched against the modules that teach it — 45 of 45 objectives have at least one module claiming them today.

1. 1.1Explain scaled dot-product attention (Query, Key, Value) and why the √dₖ scale keeps softmax from saturating.

   Covered by M1 · LLM Architecture — 5 of 5 lessons authored.

2. 1.2Explain multi-head attention, positional encoding, and layer normalization as the supporting mechanisms around self-attention.

   Covered by M1 · LLM Architecture — 5 of 5 lessons authored.

3. 1.3Match encoder-only, decoder-only, and encoder-decoder architecture families to the training objective and task they fit.

   Covered by M1 · LLM Architecture — 5 of 5 lessons authored.

4. 1.4Extract and compare embeddings from both encoder and decoder models using cosine similarity.

   Covered by M1 · LLM Architecture — 5 of 5 lessons authored.

5. 2.1Apply zero-shot, one-shot, and few-shot in-context learning without any gradient update.

   Covered by M2 · Prompt Engineering — 5 of 5 lessons authored.

6. 2.2Apply chain-of-thought prompting and prompt templates, and recognize when the added tokens and latency are not worth the gain.

   Covered by M2 · Prompt Engineering — 5 of 5 lessons authored.

7. 2.3Explain causal language modeling as the training objective that makes autoregressive generation possible.

   Covered by M2 · Prompt Engineering — 5 of 5 lessons authored.

8. 2.4Apply constrained decoding and output-validation wrappers as decoding-time controls that do not change model weights.

   Covered by M2 · Prompt Engineering — 5 of 5 lessons authored.

9. 2.5Choose between prompting, RAG, and fine-tuning for a stated constraint on data, compute, and knowledge freshness.

   Covered by M2 · Prompt Engineering — 5 of 5 lessons authored.

10. 3.1Clean and curate a training dataset, including deduplication and class-imbalance handling, using NeMo Curator and RAPIDS cuDF.

    Covered by M3 · Data Preparation — 5 of 5 lessons authored.

11. 3.2Organize and format datasets correctly for pretraining, fine-tuning, and RAG without cross-split leakage.

    Covered by M3 · Data Preparation — 5 of 5 lessons authored.

12. 3.3Distinguish BPE from WordPiece subword tokenization by their merge rule.

    Covered by M3 · Data Preparation — 5 of 5 lessons authored.

13. 3.4Explain the vocabulary-size tradeoff and why perplexity is only comparable across models sharing one tokenizer.

    Covered by M3 · Data Preparation — 5 of 5 lessons authored.

14. 3.5Run the five-step exploratory-data-analysis checklist before fine-tuning: distribution, length, vocabulary, label, and quality.

    Covered by M3 · Data Preparation — 5 of 5 lessons authored.

15. 4.1Distinguish PTQ, QAT, and GPTQ by what each needs (retraining, calibration data, Hessian information) and what each recovers.

    Covered by M4 · Model Optimization — 6 of 6 lessons authored.

16. 4.2Quantify a knowledge-distillation tradeoff using the DistilBERT precedent: size, speed, and retained performance.

    Covered by M4 · Model Optimization — 6 of 6 lessons authored.

17. 4.3Distinguish unstructured pruning from structured 2:4 sparsity by hardware-acceleration compatibility.

    Covered by M4 · Model Optimization — 6 of 6 lessons authored.

18. 4.4Explain KV caching as the primary latency lever for autoregressive inference, and that it spends memory to buy speed.

    Covered by M4 · Model Optimization — 6 of 6 lessons authored.

19. 4.5Explain sliding-window attention and TensorRT runtime optimization as distinct levers from Triton serving.

    Covered by M4 · Model Optimization — 6 of 6 lessons authored.

20. 4.6Explain masked language modeling as the training objective for encoder foundation models.

    Covered by M4 · Model Optimization — 6 of 6 lessons authored.

21. 4.7Select and justify an optimization lever (quantization, distillation, or pruning) under a stated memory or latency constraint with a measured tradeoff.

    Covered by M4 · Model Optimization — 6 of 6 lessons authored.

22. 5.1Distinguish LoRA and other parameter-efficient fine-tuning methods from full fine-tuning by what they train and what they cost at inference.

    Covered by M5 · Fine-Tuning — 5 of 5 lessons authored.

23. 5.2Distinguish SFT, RLHF, DPO, and GRPO by which of a reward model and a critic each method needs.

    Covered by M5 · Fine-Tuning — 5 of 5 lessons authored.

24. 5.3Explain contrastive loss as the mechanism training retrieval and semantic-search embedding models.

    Covered by M5 · Fine-Tuning — 5 of 5 lessons authored.

25. 5.4Apply early stopping and measure fine-tuning impact against a stated before/after baseline.

    Covered by M5 · Fine-Tuning — 5 of 5 lessons authored.

26. 5.5Decide when to fine-tune versus prompt or use RAG, given a stated skill, data, or compute constraint.

    Covered by M5 · Fine-Tuning — 5 of 5 lessons authored.

27. 6.1Explain why perplexity is undefined for masked language models and only comparable under a shared tokenizer.

    Covered by M6 · Evaluation — 5 of 5 lessons authored.

28. 6.2Distinguish BLEU, ROUGE, and METEOR by what each was built to measure.

    Covered by M6 · Evaluation — 5 of 5 lessons authored.

29. 6.3Combine LLM-as-a-judge scoring with human review and error analysis to catch fluent-but-wrong outputs.

    Covered by M6 · Evaluation — 5 of 5 lessons authored.

30. 6.4Distinguish RAG faithfulness, answer relevancy, context precision, and context recall by which pipeline stage each scores.

    Covered by M6 · Evaluation — 5 of 5 lessons authored.

31. 7.1Match a parallelism family (data, tensor, pipeline, sequence, context, expert) to the constraint it relieves when scaling a training run.

    Covered by M7 · GPU Acceleration and Optimization — 6 of 6 lessons authored.

32. 7.2Distinguish tensor parallelism (intra-layer) from pipeline parallelism (inter-layer) as the domain's most-tested distractor pair.

    Covered by M7 · GPU Acceleration and Optimization — 6 of 6 lessons authored.

33. 7.3Explain FSDP and ZeRO as memory-sharding techniques layered on data parallelism, not a separate parallelism axis.

    Covered by M7 · GPU Acceleration and Optimization — 6 of 6 lessons authored.

34. 7.4Apply mixed-precision training and gradient accumulation to fit a training run under a memory ceiling, and use Nsight to find the real bottleneck before changing them.

    Covered by M7 · GPU Acceleration and Optimization — 6 of 6 lessons authored.

35. 8.1Distinguish dynamic batching from sequence batching in Dynamo-Triton by whether the served model is stateless or stateful.

    Covered by M8 · Model Deployment — 5 of 5 lessons authored.

36. 8.2Explain NVIDIA NIM as a prepackaged, vLLM-backed microservice layered above the general Triton serving stack.

    Covered by M8 · Model Deployment — 5 of 5 lessons authored.

37. 8.3Apply concurrent model execution, instance groups, and Multi-Instance GPU to serve multiple models or copies on shared hardware.

    Covered by M8 · Model Deployment — 5 of 5 lessons authored.

38. 9.1Read latency percentiles (p95, p99), throughput, and error rate to catch tail problems a mean average hides.

    Covered by M9 · Production Monitoring and Reliability — 5 of 5 lessons authored.

39. 9.2Distinguish data drift from concept drift, and explain why an evaluation at launch is not a standing guarantee.

    Covered by M9 · Production Monitoring and Reliability — 5 of 5 lessons authored.

40. 9.3Run regression benchmarking against prior versions as a continuous check, not a one-time launch gate.

    Covered by M9 · Production Monitoring and Reliability — 5 of 5 lessons authored.

41. 9.4Version a model and its retraining pipeline so a rollback is traceable and safe.

    Covered by M9 · Production Monitoring and Reliability — 5 of 5 lessons authored.

42. 9.5Distinguish evaluation (pre-deployment quality) from production monitoring (live operational health and drift).

    Covered by M9 · Production Monitoring and Reliability — 5 of 5 lessons authored.

43. 10.1Distinguish topical, safety/content, and security (anti-injection) guardrail rails as three distinct jobs.

    Covered by M10 · Safety, Ethics, and Compliance — 5 of 5 lessons authored.

44. 10.2Measure guardrail effectiveness against its latency and false-positive-rate cost, not coverage alone.

    Covered by M10 · Safety, Ethics, and Compliance — 5 of 5 lessons authored.

45. 10.3Explain why disaggregated, per-group evaluation is required to detect bias that an aggregate accuracy number can mask.

    Covered by M10 · Safety, Ethics, and Compliance — 5 of 5 lessons authored.

Back to the NCP-GENL prep course.