Dataset

What is a dataset?

It’s the raw material of machine learning. A model doesn’t learn from rules but from examples, and a dataset is those examples gathered in a consistent format. It could be a spreadsheet with thousands of sales, a folder of call recordings with their transcripts or millions of pages of text.

For large language models, the training dataset is vast and draws on public text from many sources. In a business project it’s usually far smaller and more specific: your categorised emails, your labelled calls, your documents.

How it works

A dataset for AI usually goes through these stages:

1. Collection. You gather examples of the real problem, with permission to use them.
2. Cleaning. You remove duplicates, errors and any personal data you don’t need.
3. Labelling. For supervised learning, each example gets its correct answer.
4. Splitting. It’s divided into three parts: training (to learn from), validation (to tune along the way) and test (to measure the final result on examples the model has never seen).

Almost everything hinges on dataset quality. If a type of case is missing, the model will handle it badly. If the data reflects past inequalities, the model may repeat them: that’s the most common source of AI bias. See also model training.

Example

A small insurance broker in Norwich wants a model to sort the claims it receives by email. It pulls together two years of messages, strips out personal details that add nothing and has two staff label each one as “home”, “motor” or “pet”.

On review, they find hardly any pet claims: the model would have almost nothing to learn from. Before training, they go looking for more examples of that type. Fixing the dataset improves the result more than any tweak to the model would.

What it means for your business

For most small businesses, the question isn’t “what dataset do I have?” but “what information about my business is written down and up to date?”. An agent running on an already-trained model doesn’t need a dataset; it needs a clear knowledge base to look things up in, which is how RAG works.

That’s how UTIO’s agents work: they answer from the documentation you load, with no training data to prepare. In the support vertical, anything missing from that documentation doesn’t get a made-up answer: the agent takes a note and opens a ticket in the CRM. And for the voice agent, recordings, transcripts and the CRM are stored in Spain.