Overview
Data Analysis has five official objectives. The first, repeated from Domain 1, is awareness of mining large datasets for insight. The second asks you to compare models with statistical performance metrics, and NVIDIA names loss functions and the proportion of explained variance as examples. The third is conducting analysis under a senior team member's supervision, which in practice means reproducible, documented work. The fourth is building graphs and charts with specialized software to convey results. The fifth is identifying relationships, trends and confounding factors that could affect research results.
For an LLM associate, the data being analyzed is usually text corpora, training logs and evaluation results rather than business tables. NVIDIA's readings for this section point to RAPIDS, cuML and text preprocessing such as stemming and lemmatization, so GPU-accelerated dataframes and basic NLP preprocessing are fair game.
How it's tested
Expect short scenarios built around a dataset, a metric table or a description of a chart. You are asked which metric shows the better model, which visualization communicates a result, or which hidden factor could explain a surprising trend. Calculations are light, but you should be comfortable reading loss and perplexity values.
-
Read the metric
Two or more models are reported with loss, perplexity, accuracy or explained variance, and you choose which performs better.
-
Choose the chart
An analysis result needs to be communicated and you select the chart type or tool that presents it clearly.
-
Spot the confounder
A correlation or improvement is reported and you identify leakage, imbalance, sampling bias or another factor that undermines it.
An analyst reports that a fine-tuned model scores 95% accuracy on a sentiment set where 94% of reviews are positive. A model that always predicts positive would score nearly the same. The class imbalance means accuracy hides poor minority-class performance, so precision, recall and F1 per class, or a confusion matrix, give the honest picture. Recognizing the skewed label distribution is the key.
Objectives 5
Official objectives from the NVIDIA outline, in outline order.
2.1 Awareness of the process of extracting insights from large datasets using data mining, data visualization, and similar techniques (also objective 1.2)Exploratory data analysis on large text and log datasets.
- What it means
- Here the focus is the analyst's workflow: load the data, profile it, summarize it and visualize it to find patterns, gaps and anomalies worth reporting.
- What to understand
- Exploratory data analysis starts with counts, summary statistics, missing values and outliers. For text it adds token-length histograms, vocabulary frequency, n-grams and duplicate detection. Stemming chops words to crude roots, while lemmatization maps them to dictionary forms. RAPIDS cuDF provides a pandas-like dataframe on GPUs, which keeps exploration fast at large scale.
- How it shows up
- You pick the exploratory step that would reveal a described quality problem in a dataset.
- Common confusion
- Confusing stemming with lemmatization; lemmatization returns real dictionary words.
2.2 Compare models using statistical performance metrics, such as loss functions or proportion of explained variance.Loss, perplexity, explained variance and classification metrics.
- What it means
- This objective covers choosing and reading the right statistic to decide which model is better for a task.
- What to understand
- Cross-entropy loss measures how well a language model predicts held-out tokens, and perplexity is its exponential, so lower is better for both. Proportion of explained variance, or R-squared, measures how much of a continuous target a regression model explains, with values nearer 1 being better. Classification uses accuracy, precision, recall and F1. Generation tasks may use BLEU or ROUGE overlap scores. Compare models only on the same data split.
- How it shows up
- A metrics table is given and you identify the stronger model or the metric that suits the task.
- Common confusion
- Thinking higher perplexity means a better language model; it means the model is more surprised by the text.
2.3 Conduct data analysis under the supervision of a senior team member.Reproducible, documented analysis work.
- What it means
- At associate level, analysis is done within a team. You run the analysis, document your assumptions and hand results to a senior reviewer who signs off on conclusions.
- What to understand
- Good practice includes versioned data, fixed random seeds, notebooks or scripts that can be rerun end to end, clearly stated filters and splits, and checks for leakage between training and evaluation data. Escalating surprising results rather than quietly adjusting them is part of working under supervision.
- How it shows up
- A scenario describes an analysis shortcut and you identify the practice that keeps it reproducible.
- Common confusion
- Treating supervision as optional sign-off; reviewers need documented, rerunnable steps to verify findings.
2.4 Create graphs, charts, or other visualizations to convey the results of data analysis using specialized software.Choosing charts and tools to communicate findings.
- What it means
- You should be able to match a finding to a visualization and a tool, such as Matplotlib, Seaborn or Plotly in Python, or TensorBoard and experiment trackers for training runs.
- What to understand
- Line charts show loss or accuracy across training steps. Histograms show distributions such as document lengths. Bar charts compare models on a metric. A heatmap displays a confusion matrix or attention weights. Scatter plots of embeddings reduced with PCA, t-SNE or UMAP reveal clusters. Label axes and units, and keep scales honest.
- How it shows up
- A result needs presenting and you choose the chart type that communicates it best.
- Common confusion
- Using a pie chart for values over time, where a line chart shows the trend.
2.5 Identify relationships and trends or any factors that could affect the results of research.Correlation, confounders, leakage and bias in findings.
- What it means
- This objective asks you to read results critically: find genuine relationships and trends, and name the factors that could make a finding misleading.
- What to understand
- Correlation does not prove causation, and a confounding variable may drive both measures. Data leakage, where evaluation examples appear in training data, inflates scores, and benchmark contamination is the LLM version of this. Class imbalance, sampling bias, distribution shift between training and production data, and small sample sizes all threaten validity.
- How it shows up
- A surprising result is described and you name the factor most likely to have distorted it.
- Common confusion
- Accepting a strong correlation as proof that one variable causes the other.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Cross-entropy loss
Average negative log-probability the model assigns to the correct tokens or labels.
- Remember
- Lower is better.
- Tested as
- Model comparison questions.
- Trap
- Reading a higher loss as a stronger model.
-
Perplexity
The exponential of cross-entropy loss for a language model.
- Remember
- Lower perplexity means better prediction of held-out text.
- Tested as
- Language model evaluation questions.
- Trap
- Comparing perplexity across different tokenizers as if equal.
-
Explained variance (R-squared)
Share of a continuous target's variance a regression model accounts for.
- Remember
- Closer to 1 is better.
- Tested as
- Regression comparison questions.
- Trap
- Applying it to classification or text generation.
-
Precision, recall and F1
Precision is correctness of positive predictions; recall is coverage of actual positives; F1 balances them.
- Remember
- Use them when classes are imbalanced.
- Tested as
- Classifier evaluation questions.
- Trap
- Trusting accuracy on skewed data.
-
Confusion matrix
A table of predicted versus actual classes.
- Remember
- Shows exactly which classes are confused.
- Tested as
- Visualization and metric questions.
- Trap
- Reading only the diagonal total.
-
Exploratory data analysis
Profiling a dataset with statistics and charts before modeling.
- Remember
- It surfaces gaps, outliers and duplicates.
- Tested as
- Data insight questions.
- Trap
- Skipping it and modeling raw data.
-
Stemming vs. lemmatization
Stemming trims suffixes crudely; lemmatization returns dictionary forms using grammar.
- Remember
- Lemmatization is more accurate, stemming faster.
- Tested as
- Text preprocessing questions.
- Trap
- Treating them as the same.
-
RAPIDS cuDF
GPU-accelerated dataframe library with a pandas-like API.
- Remember
- It speeds analysis of very large tables.
- Tested as
- NVIDIA tooling questions.
- Trap
- Thinking it is a model training framework.
-
Dimensionality reduction for plots
PCA, t-SNE or UMAP project embeddings to 2D for display.
- Remember
- Useful for seeing clusters, not exact distances.
- Tested as
- Embedding visualization questions.
- Trap
- Reading t-SNE distances literally.
-
Data leakage
Information from evaluation data reaches training or tuning.
- Remember
- It inflates reported performance.
- Tested as
- Validity questions.
- Trap
- Missing duplicates across splits.
-
Confounding variable
A third factor influences both measured variables.
- Remember
- It can create a misleading correlation.
- Tested as
- Research factor questions.
- Trap
- Concluding causation from correlation.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Model comparison metrics
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Cross-entropy loss | Penalty for low probability on correct outputs | Lower is better, training and evaluation | Comparing across different datasets |
| Perplexity | Exponential of language model loss | Lower is better, intuitive scale | Comparing models with different tokenizers |
| R-squared | Proportion of explained variance | Regression only, higher is better | Using it for classification |
| F1 score | Harmonic mean of precision and recall | Robust to class imbalance | Ignoring per-class results |
Choosing a chart
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Line chart | Values across ordered steps or time | Shows trends such as loss curves | Using bars for continuous training curves |
| Histogram | Distribution of one variable | Shows spread and skew | Confusing it with a bar chart of categories |
| Heatmap | Color-coded matrix values | Suits confusion matrices and correlations | Omitting a color scale |
| Scatter plot | Two variables or reduced embeddings | Reveals clusters and relationships | Overplotting without sampling |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Higher perplexity wins
- What candidates pick
- Pick the model with the larger perplexity.
- Why it looks right
- Bigger numbers often mean better scores.
- Why it's wrong
- Perplexity measures surprise, so the lower value indicates better prediction of the text.
Clue to watch for The table reports perplexity or loss.
-
Accuracy on skewed data
- What candidates pick
- Praise a high accuracy score on an imbalanced set.
- Why it looks right
- Accuracy is the familiar headline metric.
- Why it's wrong
- A model predicting only the majority class can score highly; per-class precision, recall and F1 reveal the gap.
Clue to watch for One class makes up most of the labels.
-
R-squared for language models
- What candidates pick
- Compare text generators by explained variance.
- Why it looks right
- The objective names explained variance.
- Why it's wrong
- Explained variance applies to continuous regression targets, not token prediction or classification.
Clue to watch for The models predict text or categories.
-
Correlation is cause
- What candidates pick
- Conclude that one factor drives another because they move together.
- Why it looks right
- The pattern looks convincing.
- Why it's wrong
- A confounder or coincidence can produce correlation without any causal link.
Clue to watch for The scenario reports an association without a controlled test.
-
Pie chart for trends
- What candidates pick
- Show training loss over epochs in a pie chart.
- Why it looks right
- Pie charts look simple.
- Why it's wrong
- Pie charts show parts of a whole; trends over steps need a line chart.
Clue to watch for The data is ordered by time or step.
Answer framework
Apply this sequence when a Domain 2 question gives you data, metrics or a chart to interpret.
- Identify the task type first, such as language modeling, classification or regression, because it determines which metric is meaningful.
- Check the direction of each metric, remembering that loss and perplexity improve downward while R-squared, F1 and accuracy improve upward.
- Look at the data behind the numbers, including class balance, split overlap and sample size, before trusting a result; high accuracy on a 95-to-5 imbalanced set can hide a model that never predicts the minority class.
- If asked to communicate a finding, match its shape to a chart: trend, distribution, comparison, matrix or cluster.
- Before accepting any relationship, name a plausible confounder, source of leakage or distribution shift that could explain it instead, and treat a correlation as a lead to test rather than proof of cause.
Practice question
Want more like this? The NCA-GENL practice test has questions for every domain.
Checklist
0 of 12
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
This is a smaller domain, so aim for complete coverage rather than depth. Start with metrics because they generate the most questions. Write a one-page sheet listing each metric, what it measures, which task it fits and whether higher or lower is better. Practice converting cross-entropy into perplexity with a calculator until it feels automatic, and work through a few imbalanced-class examples to see accuracy mislead.
Next, open a notebook with a public text dataset. Load it with pandas or cuDF, plot token lengths, count duplicates, and compare stemmed and lemmatized vocabularies. Then plot a loss curve, a confusion matrix heatmap and a 2D projection of some embeddings. Making each chart once fixes the chart-to-purpose mapping in memory.
Finish with the critical-thinking objective. For every result you see in practice questions, ask what could have produced it other than the claimed cause: leakage, contamination, imbalance, shift or a confounder. That habit carries straight into Experimentation and Trustworthy AI. Plan about a week of part-time study, then confirm with timed questions.
FAQ
How much of NCA-GENL is the Data Analysis domain?
Data Analysis is 14% of the exam, which is roughly seven to nine questions on a 50 to 60 question form.
Is perplexity better when it is higher or lower?
Lower. Perplexity is the exponential of cross-entropy loss, so a lower value means the model is less surprised by real text.
What is proportion of explained variance?
It is R-squared, the share of a continuous target's variation that a regression model explains. Values near 1 indicate a better fit.
Which visualization tools should I know?
Know common Python plotting libraries such as Matplotlib, Seaborn and Plotly, and training dashboards such as TensorBoard. The exam focuses on choosing a suitable chart more than on syntax.
Where do RAPIDS tools fit in?
RAPIDS libraries such as cuDF and cuML run dataframe operations and classic machine learning on NVIDIA GPUs, which NVIDIA lists in this section's readings.
Is the official domain called Data Analysis and Visualization?
NVIDIA's exam page and study guide name it simply Data Analysis, though one objective covers creating graphs and charts.
How does this domain connect to Experimentation?
Experimentation produces runs and benchmark scores; Data Analysis is how you compare them honestly and explain what drove the differences.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 2 knowledge
You have covered data exploration, model metrics, visualization and threats to valid results. See whether you can read a metrics table and pick the right chart quickly.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the NCA-GENL AI tutor Get any concept explained your way