NCP-GENL Fine-Tuning Practice Question
A team fine-tunes a model with NVIDIA NeMo using a packed sequence dataset and notices that some training samples contain several short conversations concatenated. They must ensure the loss is computed only on assistant responses and not on the packed boundaries. Which configuration detail should they verify?
⚠ Common exam trap
The trap here is treating packed sequences as a pure throughput optimization and overlooking that masks and attention boundaries must be adjusted to keep the loss correct.
Answer choices
Why each option matters
Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.
Correct answer & explanation
✓
That the loss mask aligns with each sample's response tokens and that packed sequences are separated by an attention boundary
Packed sequences boost training efficiency by filling each context window with multiple samples, but correctness depends on the loss mask covering only assistant response tokens and on attention being blocked across sample boundaries. Batch size, vocabulary changes, and gradient accumulation do not affect these two requirements.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
That the global batch size is increased to compensate for the additional tokens in each packed sequence
Why it's wrong here
Batch size affects throughput and optimization dynamics but has no bearing on whether the loss mask or attention boundaries are correct. Increasing batch size would not prevent the model from training on tokens outside the intended responses. This is a performance tuning knob, not a correctness fix for packed data.
- ✗
That the tokenizer vocabulary is expanded to include special separators between packed samples
Why it's wrong here
Adding vocabulary entries does not by itself create attention isolation between packed samples or mark response tokens for the loss. NeMo handles packing boundaries through sequence identifiers and masks rather than new vocabulary tokens. Expanding the vocabulary would also change embedding shapes and require retraining.
- ✓
That the loss mask aligns with each sample's response tokens and that packed sequences are separated by an attention boundary
Why this is correct
Packed sequences concatenate multiple samples into one training sequence for efficiency, so the loss mask must mark only assistant response tokens and the attention mechanism must prevent cross-sample attention. In NeMo this is handled through per-token loss masks and sequence boundary handling. Verifying both ensures the model is not trained on padding or on tokens from adjacent samples.
- ✗
That gradient accumulation steps equal the number of samples packed into each sequence
Why it's wrong here
Gradient accumulation controls how many micro-batches contribute to one optimizer step and is unrelated to how loss masks or attention boundaries are applied within a packed sequence. Tying it to the packing count has no technical basis and would distort the effective batch size. It does not address the correctness concern.
About these practice questions
This NCP-GENL question is part of Courseiva's 352-question bank — original exam-style content with full explanations and wrong-answer analysis, never real exam questions or exam dumps. Learn why practice questions differ from exam dumps →
JA
Written and reviewed by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
Last reviewed September 2026 · checked against the official NVIDIA exam blueprint
This NCP-GENL practice question is part of Courseiva's free NVIDIA certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the NCP-GENL exam.