NCA-GENL Experimentation Practice Question
A data science team is running a controlled experiment with NVIDIA NeMo to compare two fine-tuning recipes for a 7B-parameter LLM: one with a constant learning rate and one with a cosine decay schedule. They notice the evaluation loss curves diverge significantly after step 500, but they cannot tell whether the difference is caused by the learning-rate schedule or by random seed variance. Which experimental change should they make to isolate the effect of the schedule?
⚠ Common exam trap
The trap here is assuming that smoother curves or more data automatically make an experiment conclusive, when the real issue is uncontrolled seed and data-order variance confounding the comparison.
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
✓
Run both recipes with a fixed random seed and identical data ordering, then compare the loss curves.
A controlled experiment requires isolating the independent variable—here the learning-rate schedule—while holding all other factors constant. Random seed and data ordering are common sources of run-to-run variance in LLM fine-tuning. Fixing them across both recipes ensures that observed differences in evaluation loss are caused by the schedule rather than initialization or batch order, making the comparison valid.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✓
Run both recipes with a fixed random seed and identical data ordering, then compare the loss curves.
Why this is correct
Holding the random seed and data ordering constant removes the confounding effect of initialization and batch-order variance, so any remaining divergence between the constant and cosine schedules can be attributed to the learning-rate schedule itself. This is the core principle of a controlled experiment: change one factor at a time while controlling all others, including seeds and data shuffling.
- ✗
Switch both recipes to a warmup-stable-decay schedule and compare the final evaluation loss instead of the curves.
Why it's wrong here
Replacing both schedules with a third schedule eliminates the original comparison entirely. The team wanted to compare constant versus cosine decay, not a new schedule. While comparing final loss can be useful, it does not address the confounding seed variance and abandons the experimental question being asked.
- ✗
Increase the batch size for both recipes until the loss curves become smoother and easier to compare visually.
Why it's wrong here
Increasing batch size changes the optimization dynamics and can alter the effective learning rate, making the two schedules even harder to compare. It also does not remove seed-driven variance; it merely changes the noise profile. This would confound the experiment further rather than isolate the learning-rate schedule as the single independent variable.
- ✗
Run each recipe on a different GPU type to see whether hardware differences explain the divergence.
Why it's wrong here
Introducing different hardware adds a new confounding variable rather than controlling one. GPU type can affect numerical precision, kernel selection, and throughput, all of which can change loss trajectories. This would make it impossible to attribute divergence to the learning-rate schedule, defeating the purpose of the controlled comparison.
About these practice questions
This NCA-GENL question is part of Courseiva's 367-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 NCA-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 NCA-GENL exam.