GlossaryKind of term: Manufacturing termTraining data
What is training data?
Training data is the set of examples from which an AI model is built or adjusted, such as text, labelled drawings or parts with measured times.
During training the model's parameters are set from those examples. They differ by kind of model: text for a language model, labelled images or drawings for computer vision, parts with measured times for a time model. In supervised learning each example carries a label, the correct answer, such as the actual time or the tolerance as it stands on the drawing. Training data differs from the input a system processes to answer one question: input becomes training data only when it is kept and used to build or adjust a model.
How is training data split?
It is common practice to split the examples into a training set, a validation set for tuning and a test set held back until the end. Performance measured on the training data itself says little. The examples have to cover the work the model will be used on, with correct labels; errors and exceptions in the labels end up in the model. An example that sits in both the training and the test set makes the result look better than it is.
How does training data work in practice?
For a shop, the main questions are rights and confidentiality. A customer drawing is not yours; confidentiality agreements usually say whom you may pass it on to. Your own hourly rates, purchase prices and cycle times are commercially sensitive. Ask the supplier of any system that works with AI, in writing, whether files are used to train models, whom that model serves and whether you can opt out. Record the answer in the agreement.
What is training data often confused with?
A drawing you upload for one calculation is input; it becomes training data only if it is kept and used for a model. Test data is different again: the test set serves to assess a trained model and should stay out of training.
Frequently asked questions
Is my data used for training?
That depends on the supplier and the agreement. Get it in writing whether drawings, times and prices go into model training, whom those models serve and whether you can opt out.
May I put a customer drawing into an AI tool?
Check first what you agreed with the customer on confidentiality. If drawings may not go to third parties, using an outside tool needs the customer's consent or an agreement with the supplier on processing, storage and training.
What should a contract say about training data?
For example, that drawings, times and prices are processed only for your work, are not used for training without written consent, and where and for how long they are kept. Have a lawyer review the wording.