--- language: en title: "Training data" description: "Training data consists of examples used to adjust a model’s parameters. Their selection, origin, and composition influence the behavior learned." canonical: "https://tropk.ai/en/resources/glossary/training-data" last_modified: "2026-10-12T16:13:33.647Z" --- # Training data Reviewed on 2026-10-06. ## What it means Training data consists of examples used to adjust a model’s parameters toward a learning objective. In language models, it can include text sequences used to predict tokens. The origin, selection, and processing of these examples make up the dataset actually used; being available on the internet does not necessarily mean that something is part of that dataset. ## Why it matters for marketing Marketing teams need to understand this distinction when choosing models or preparing [fine-tuning](/en/resources/glossary/fine-tuning) with their own materials. A dataset dominated by another language or context may require additional evaluation for Brazilian tasks. For example, the GPT-2 documentation describes WebText and acknowledges biases associated with the data. This history helps frame questions about composition without describing all current models. ## Practical example Suppose an agency prepares training data to fine-tune a model for writing regional campaigns. When reviewing the sample, it finds only ads from São Paulo, although the application also serves the Northeast. Before fine-tuning, the team expands the examples to include regional language and offers, and sets aside separate materials to evaluate whether the learned behavior meets the intended scope. ## How to apply or verify Request documentation on the dataset, languages, time period, selection criteria, and transformations performed. In a project of your own, review samples and keep evaluation examples separate from those used for training. Also check the authorized purpose for the materials and whether they contain confidential information. If the provider does not disclose the full dataset, record this limitation rather than inferring its composition. ## Important distinctions Training data differs from documents retrieved for a response and from the brief sent in a call: these provide context without necessarily changing parameters. It is also important to distinguish pre-training data from data used for later fine-tuning. A response that mentions a brand does not prove that a specific page was included in training, nor does it identify when it was included. ## Sources and references - [GPT-2 model card]() — OpenAI - [GPT-2]() — Hugging Face ## Continue a leitura - [Training data cutoff]() - [Fine-tuning]() - [Parametric memory]() ## About the author [Vinicius Pacheco]() Responsible for Tropk’s technical content, focusing on crawling, indexing, structured data, and agent access to content. --- [All pages in Markdown]()