Training data is the set of example inputs, usually paired with correct outputs, that a machine learning model learns from by adjusting its internal parameters to fit them. In commercial real estate extraction, it is a collection of leases, rent rolls, or offering memorandums whose fields are already marked correct, teaching a model what base rent looks like.
What Is Training Data in Machine Learning?
Training data in machine learning is the labeled subset a model learns from, where each input carries the correct output so the model can map one to the other. The model adjusts its parameters to fit these examples, per Statology's overview of train-test splits. Quality and coverage of this data set the ceiling on how well the model performs on documents it has never seen.
Training data is only part of the picture. A dataset is split so the model is trained on one portion and judged on another it never touched. A common split is 80% training and 20% test, sometimes 70/30, with the larger share for training so the model has enough to learn from while leaving an honest test, per Statology.
Term | Role |
Training set | The portion the model learns from |
Test set | Held-out data used to judge performance honestly |
Label | The correct output attached to each training example |
Stratified split | Split that preserves label balance across both sets |
Why Training Data Matters
Training data matters because a model can only recognize what its examples taught it. Feed it 500 office leases and no retail leases, and it will stumble on a percentage rent clause it never saw. Coverage of formats, asset types, and edge cases in the training data is what decides whether extraction holds up across a real portfolio.
Scale illustrates the point. OpenAI's GPT-3, with 175 billion parameters, was trained on roughly 300 billion tokens of text, per NVIDIA's reporting. That breadth is why a large language model generalizes across lease phrasings. For a narrow classifier, the lesson is the same at smaller scale: gaps in the training data become blind spots in production.
Example
A team builds a model to flag renewal option clauses and assembles 800 leases where a human has already marked each clause. They split the set 80/20: 640 leases for training and 160 held out for testing. The model learns only from the 640 and is scored on the 160 it never saw.
Component | Figure |
Total labeled leases | 800 |
Training set (80%) | 800 x 0.80 = 640 leases |
Test set (20%) | 800 x 0.20 = 160 leases |
Retail leases in set | 0 |
The test score looks strong, but every one of the 800 leases is office. When a retail lease arrives with a co-tenancy clause the model never trained on, it misses. The fix is not a better algorithm; it is training data that includes retail. The example shows a rule operators repeat: a model is only as good as the range of documents in its training data.
Variations and Edge Cases
Training data varies by how it is labeled, balanced, and split. Each choice below changes what the model can learn and how honestly its score reflects real performance.
Variant | Treatment |
Labeled data | Examples with correct outputs, needed for supervised learning |
Unlabeled data | Raw documents with no marked answers |
Validation set | A third split used to tune settings before the final test |
Imbalanced data | One label dominates, skewing a naive model |
Data leakage | Test information seeps into training, inflating the score |
Training Data vs Test Data
Training data is often confused with test data, and the difference is purpose. Training data is what the model learns from, adjusting its parameters to fit those examples. Test data is held out and never seen during training, used only to measure how well the model performs on new documents.
Mixing the two is the classic error. If a lease appears in both sets, the model can memorize it and post a score it cannot repeat in production, a failure called data leakage. Keeping training and test data strictly separate is what makes an accuracy figure trustworthy before a model touches a live deal.
Frequently Asked Questions
What is training data in simple terms?Training data is the set of example inputs, usually with correct answers attached, that a model learns from. In commercial real estate it is a batch of leases or rent rolls whose fields are already marked correct, teaching the model what base rent, term, or occupancy looks like.
How much training data does a model need?It depends on the task. A large language model like GPT-3 trained on roughly 300 billion tokens, per NVIDIA, while a narrow lease classifier may learn from a few hundred labeled documents. What matters more than raw volume is coverage of the formats and edge cases the model will face in production.
What is the difference between training data and test data?Training data is what the model learns from; test data is held out and used only to score the model on documents it never saw. Keeping them separate prevents data leakage, where a model memorizes examples and posts an accuracy figure it cannot repeat on new deals.
Related Terms
Machine Learning Model
Large Language Model
Field Extraction
Confidence Score
Structured Data Extraction