Glossary
What is AI training data?
In short
AI training data is the set of examples on which a model learns and with which it later works: texts, tables, photos, call recordings and sales history.
AI basics · training data
MMOOZZGG.com editorsChecked: 1 min read
For a child
To learn to make pancakes, you need to try many times and watch how grandma does it. A computer also needs examples to learn: pictures, words and numbers. If the examples are clear and varied, it learns well. If all the pictures are dark and blurry, it gets confused. So people carefully collect and check what they show the computer.
In plain words
Any trained system knows only what was in its examples. If you show it photos of only ginger cats, it may not recognize a gray one. So people collect varied examples and often label them by hand: “this photo shows a cat”, “this letter is a complaint”. Besides training, AI needs working materials: an assistant for contracts needs the company’s contracts. How complete and clean these materials are often affects the answers as much as the choice of algorithm.
For business
For a business this is usually not new information but what already sits in the CRM, 1C accounting software, email, call recordings and spreadsheets. The trouble is that it is scattered across systems, with gaps and duplicates. Preparation, meaning export, cleaning, labeling and description, often takes longer than setting up the model itself. In Russia, client details fall under the law “On Personal Data” No. 152-FZ of 27 July 2006, so it matters where they are stored and who receives them. We describe on a separate page what to measure in processes, how to collect it and where to send it.
For a specialist
Training data is split into training, validation and test sets, so that quality is measured fairly on examples the model has not seen. Supervised learning needs labels, and their reliability is checked by inter-annotator agreement. Typical problems are data leakage from the test set into training, sampling bias, class imbalance and data drift after launch. For large language models, deduplication, quality filtering and source licenses also matter. In 2018, T. Gebru and co-authors proposed that every dataset come with a datasheet: its purpose, composition, collection process, recommended uses and limits.
Example
A clinic network wants AI to sort patient requests. First, analysts export past years’ requests from the CRM, remove names and phone numbers and mark the topic of each one by hand.
Related terms
Questions and answers
Where can a company get training data?
Sources
Not sure whether your company needs this?
AI solutions map