AI Data Lab Wiki

AI Data Lab Wiki › Glossary › AI training data

What Is AI Training Data?

By AI Data Lab Wiki editors · checked November 15, 2026

AI training data is the information used to teach a machine learning model to make predictions, recognize patterns or generate content. IBM defines it that way and adds that without it even sophisticated algorithms are not useful. 1

What Is AI Training Data

AI training data is a dataset the model processes so it can find patterns, and a dataset can be made up of spreadsheets, video footage, web pages, PDFs or other data. 1

IBM says the amount of data matters, and so does its quality. Poor quality data can introduce noise and bias that keep a model from making accurate predictions.

Types of AI Training Data

AI training data comes in labeled and unlabeled forms, and IBM says the type depends on the learning method: supervised learning uses labeled data, unsupervised learning uses unlabeled data, and semi-supervised learning combines both. 1

IBM's example for labeled data is spam detection. Thousands of emails are marked by humans as spam or not spam, and the model learns which patterns go with each mark. The marking step is covered under data annotation.

How AI Training Data Is Prepared

AI training data is prepared in stages, and IBM describes the work as collecting, cleaning, transforming and splitting the data. 1

  1. Collect the raw data from sensors, public datasets, clickstreams or other sources.
  2. Clean it by fixing missing values, duplicates and other errors.
  3. Transform it into a format an algorithm can process.
  4. Label it, where the method needs labels, often with human annotators who may need domain expertise such as law.
  5. Divide it into a training set, a validation set and a testing set, so the model can be checked on data it has not seen.

IBM notes that human review in this process is sometimes called human in the loop. See human-in-the-loop for the term.

LLM Training Data

LLM training data usually has two layers: a very large collection of text, and smaller sets of human-written examples that shape how the model behaves.

For the first layer, Common Crawl publishes a corpus of raw web page data, metadata and text extracts that it says holds petabytes of data collected since 2008. Common Crawl is one public source of web text. 2

The GPT-3 paper reports a model with 175 billion parameters and says earlier approaches needed task-specific fine-tuning datasets of thousands or tens of thousands of examples. 3

For the second layer, the InstructGPT paper describes a dataset of labeler demonstrations, used for supervised fine-tuning, and a dataset of rankings of model outputs, used with reinforcement learning from human feedback. The authors reported that outputs from a 1.3 billion parameter InstructGPT model were preferred to outputs from the 175 billion parameter GPT-3. 4

AI Datasets

AI datasets are collections of examples that can be documented, and a 2018 paper called Datasheets for Datasets proposed that every dataset come with a datasheet describing its motivation, composition, collection process and recommended uses. 5

The authors said machine learning had no standard process for documenting datasets and argued that a datasheet would improve communication between dataset creators and users.

Who Makes AI Training Data

AI training data is made by AI companies themselves, by public projects and by AI data labs that sell human-made data to model builders. IBM says many businesses hire large teams of human data annotators, sometimes assisted by machines. 1

The wiki's list of AI data labs covers the companies that sell this work, and the AI trainer page covers the job. Some data is also produced by software, which is covered under synthetic data.

Questions About AI training data

These are the questions about AI training data that people search most.

What is AI training data used for?

AI training data is used to teach a model to predict outcomes, recognize patterns or generate content. After training, a separate testing set is used to check how well the model handles data it has not seen.

What is LLM training data?

LLM training data is the text and human-written examples used to build a large language model. It includes large web collections such as Common Crawl, plus smaller sets of labeler demonstrations and ranked outputs.

What are AI datasets?

AI datasets are organized collections of data used to train or test models, such as spreadsheets, video, web pages or PDFs. A 2018 paper proposed documenting each one with a datasheet covering how it was collected.

Does AI training data need labels?

AI training data needs labels only for some methods. IBM says supervised learning uses labeled data, unsupervised learning works with unlabeled data, and semi-supervised learning combines both kinds.

More AI Data Terms

More AI data terms explained in this wiki:

Sources for AI training data

The sources for AI training data are listed in the order the page cites them, read November 15, 2026.

  1. ibm.com/think/topics/training-data
  2. commoncrawl.org/overview
  3. arxiv.org/abs/2005.14165
  4. arxiv.org/abs/2203.02155
  5. arxiv.org/abs/1803.09010