Before training, you find that 90% of your labels are one class. What is the risk?
Why C is the answer
A model that looks accurate while ignoring the minority class
With imbalance, predicting the majority class alone gives high accuracy, so use better metrics and rebalance.
In detail
- A model can score well by always choosing the majority class.
- Use precision, recall or F1 here.
Use case
An illustrative scenario, not a real incident.
A model that always says "not fraud" scores 90% on skewed labels, and catches nothing.
See it in different ways
flowchart LR A["Dataset"] --> B["Training: fit"] A --> C["Validation: tune"] A --> D["Test: final score"]
mindmap
root((ML basics))
Labeled data
Classification
Regression
Unlabeled data
Clustering
Validation
Cross validation
Held-out test setflowchart TD
Q{"Is one class dominant?"}
Q -->|"Yes"| A0["Rebalance and use better metrics"]
Q -->|"No"| A1["Accuracy may be fine"]Key point: Accuracy misleads on imbalanced data.
Why the other options are wrong
- A. Too many embeddings for the chosen vector dimension
- Dimension is not affected.
- B. GPU memory errors caused by uneven batches
- Class balance does not cause memory errors.
- D. Slower tokenization on the majority class
- Tokenization speed is not affected.