What is model training

Definition

Model training is how an AI model learns. It is shown a large number of examples, makes a guess for each, compares the guess with the right answer and adjusts its internal settings to be a little less wrong, round after round, until the errors stop shrinking. Large language models are trained first on vast amounts of text, learning to predict the next piece of it. The chat tool you use was trained this way by its vendor, on text up to a date, so it knows nothing after that date and nothing about your business unless you give it the information. Uploading files to a chat tool is not training: the tool searches the files, and the model stays the same.

Why it matters

Your part is different. Give the model your facts when it answers, and know whether your own conversations may be used to train the vendor's next version, which is a setting of your plan.

Directly related: Inference, AI Model, Fine-tuning.

Improvement tips

• Check your plan's setting on whether your conversations may be used to train future models, and choose on purpose.
• Give the model your business facts in its instructions or documents, since its training did not include them.
• Judge any trained model on new examples it never saw, not on the ones it learned from.
• For anything recent, give the model the source or use a tool that searches, because its knowledge stops at a date.

Common mistakes

• Judging a model by its score on the examples it was trained on.
• Assuming a trained model knows recent events or your current hours.
• Leaving the plan's training setting at its default without knowing what the default is.