Say it like this
- We need more training data to improve the model accuracy
- The quality of our training data directly affects model performance
- We cleaned up the training data before feeding it to the algorithm
Use it at work
ML
- We need more training data for rare cases.
- The training data is biased toward English.
Privacy
- Can we use customer data as training data?
- We removed personal info from the training data.
Debugging
- The model is wrong because the training data had bad labels.
- Let's check the training data.
Don't say this
Training data is for teaching the model; test data is for evaluating its performance on unseen examples.
Don't mix it up
| Term | What it is | Who writes or uses it | When |
|---|---|---|---|
| training data | Data used to teach the model | ML engineers | Training |
| test data | Data held back to check the model | ML engineers | Evaluation |
| ground truth | Correct answers to compare against | ML engineers | Evaluation |
Practice
Try it
4 quick exercises
1. Your manager asks about the dataset for your machine learning project. What do you say?
2. Before we train the model, we need to _______ the training data to remove errors and inconsistencies.
3. Which sentence uses training data incorrectly?
4. Rewrite this in natural tech English using training data: The information we use to teach the computer program to recognize patterns needs to be accurate and come from many different sources.
Say it out loud
Phrase 1 of 3
"We need more training data to improve the model accuracy"
We don't save your recording. The word check uses your browser's speech recognition.
More about training data
Real examples(3)
- in a standup
"We collected 50,000 images for training data this sprint to improve our object detection model."
- in a code review
"The training data we used includes samples from all geographic regions to prevent regional bias."
- in a team meeting
"Our labeling team is still processing training data, so we won't be ready to train the model until next week."
More mistakes(2)
Handle with care. Do not confuse with test data or validation data. Training data is specifically used to teach the model, not to evaluate it.
- ✕The training data is invalid because the model performed poorly.✓The model performed poorly, which suggests we need to review our training data quality and diversity.
Poor performance does not automatically mean the training data is invalid; it requires investigation into multiple factors.
- ✕We used all our training data for testing.✓We split our data into training and test sets to avoid overfitting.
Using training data for testing creates inflated performance metrics and does not reflect real performance.
Other forms(3)
- training data noun
"The training data consists of labeled examples that the model learns from."
- training dataset noun
"Our training dataset contains 100,000 labeled customer reviews."
- training adjective
"The training process requires clean and diverse data."
Often used with(5)
- high-quality training data · Emphasizes that the data is clean, accurate, and well-labeled.
- training data augmentation · Techniques to artificially expand the training dataset by creating variations of existing examples.
- labeled training data · Training data where each example has been annotated with the correct answer or label.
- training data bias · Systematic errors or skews in the training data that cause the model to learn unfair patterns.
- prepare training data · The process of cleaning, organizing, and formatting data before it is used to train a model.
Similar words(3)
- training set · Slightly more formal term; often used interchangeably with training data.
- training dataset · More explicit and specific; emphasizes that it is a complete collection of examples.
- labeled data · Broader term that can refer to any annotated data; not specific to training.
Opposites(3)
- test data
- validation data
- unlabeled data
Questions(5)
- What is the difference between training data and test data?
- Training data is used to teach the model by showing it examples and correct answers. Test data is used to evaluate how well the model performs on data it has never seen before. Using training data to test your model gives misleading results because the model has already learned from those examples.
- How much training data do I need?
- The amount varies depending on your problem complexity and the model type. Simple problems may need thousands of examples, while complex problems may need millions. Generally, more diverse and high-quality data is better than simply having more data. It is common to start small and add more data if model performance is poor.
- What happens if my training data is biased?
- If your training data is biased, your model will learn and perpetuate those biases. For example, if your training data contains mostly images of one demographic group, your model will perform poorly on other groups. This is why teams carefully audit training data for fairness and diversity.
- Can I use the same data for both training and testing?
- No, you should never use the same data for both training and testing. This creates overfitting and gives false confidence in your model. Best practice is to split your data into separate training, validation, and test sets.
- What is data labeling in the context of training data?
- Data labeling is the process of annotating or tagging examples in your training data with the correct answers. For instance, in an image classification task, you would label each image with the object it contains. Labeled data is essential for supervised learning.
