What is training data?

Training data is the information an AI model learns from, such as text, images or past business records. Large language models are trained on huge amounts of public text, including websites, books, research papers and code. On its own, a model knows only what was in its training data, up to the date the data was collected.

Why it matters for a business

What is in the training data decides what a model is good at, how it sounds and where it fails. A company’s internal documents, customers and prices were never in a public model’s training data, so the model answers questions about them by guessing from businesses that look similar. The data also has an end date, the knowledge cutoff. And whatever patterns the data holds, including unfair ones inside past decisions, the model learns and repeats.

Go deeper

How AI learns: the 4 stages of training a model

Related terms

• Knowledge cutoff
• AI bias
• Fine-tuning
• RLHF (reinforcement learning from human feedback)
• Machine learning

All terms, A to Z