Training data

Think of training data like the textbooks and practice problems a student works through before an exam. The student doesn’t memorize every answer—they learn underlying patterns, relationships, and problem-solving strategies that they can apply to new questions they’ve never seen before. Training data plays the same role for an AI model: it provides the examples from which the model learns the patterns it will later generalize from.

How models learn from training data

During training, a machine learning model processes training data iteratively, adjusting its internal parameters (weights and biases) to minimize the difference between its predictions and the actual outcomes in the data. In large language models (LLMs), training typically happens in two phases:

Pre-training is the initial phase where the model learns from a massive, general-purpose dataset. For LLMs, this often means billions of text passages drawn from books, websites, academic papers, and other written sources. The model learns to predict the next token in a sequence, developing a broad understanding of language, facts, and reasoning patterns along the way.

Fine-tuning is a later, more focused phase where a pre-trained model is further trained on a smaller, task-specific dataset. For example, a model pre-trained on general text might be fine-tuned on medical records to perform healthcare tasks, or on customer support transcripts to handle service inquiries. Fine-tuning adapts the model’s general capabilities to a specific domain or behavior.

Types of training data

Training data varies widely depending on the learning approach:

• Labeled data includes both the input and the correct output. For example, a dataset of emails where each is tagged as “spam” or “not spam” is labeled data. Models trained on labeled data learn to map inputs to known outputs—this is called supervised learning.
• Unlabeled data includes only the input, without a corresponding correct answer. A large collection of raw text without annotations is unlabeled data. Models can still learn useful patterns from unlabeled data through techniques like clustering or self-supervised learning.
• Human feedback data captures human preferences or corrections, often used in reinforcement learning from human feedback (RLHF). This type of data helps align model behavior with human values and expectations—critical for making AI systems that are helpful, harmless, and honest.
• Synthetic data is artificially generated, often by another AI model, to supplement real-world data or create training examples that are rare or expensive to collect. Synthetic data is increasingly used in scenarios where real data is scarce, expensive, or privacy-sensitive.

Why data quality matters

The quality of training data directly shapes what a model can learn. Several factors affect training data quality:

Representativeness: Training data should reflect the diversity of real-world scenarios the model will encounter. If a facial recognition model is trained primarily on images of one demographic, it will perform poorly on others—a well-documented example of how skewed training data creates biased models.

Accuracy: Errors in training data—mislabeled examples, incorrect facts, or corrupted entries—can teach the model wrong patterns. This is particularly problematic for LLMs, which may hallucinate confidently if trained on inaccurate source material.

Bias: Historical biases present in training data can be absorbed and amplified by models. If a hiring model is trained on historical data that reflects discriminatory practices, it may learn to replicate those biases without anyone explicitly teaching it to discriminate.

Freshness: Models trained on outdated data may not reflect current realities. A model trained on pre-2020 medical literature won’t know about treatments developed since then. This is why some organizations periodically retrain models on updated datasets.

Security and governance concerns

Training data is also a security surface. Data poisoning attacks involve deliberately inserting malicious or misleading examples into training data to compromise model behavior. The OWASP Top 10 for LLM Applications (2025 edition) identifies data poisoning as a critical supply chain risk, noting that manipulated training data can introduce vulnerabilities, backdoors, or biases that are difficult to detect after training is complete.

NIST AI 600-1, the U.S. government’s guidance on generative AI risk management, emphasizes that training data governance—including documentation of data sources, provenance, licenses, and collection methods—is essential for building trustworthy AI systems. Organizations are increasingly expected to maintain detailed records of what data was used to train their models and how that data was collected and processed.

Training data vs. other datasets

A complete machine learning pipeline typically uses three distinct datasets:

• Training data: Used to teach the model—this is where the actual learning happens.
• Validation data: Used during training to check whether the model is overfitting (learning the training data too specifically rather than learning general patterns).
• Test data: Used after training to evaluate how well the model generalizes to truly unseen examples.

These three datasets should never overlap. Using the same data for training and testing would give an artificially optimistic assessment of model performance, since the model would essentially be tested on material it has already memorized.

Why it matters for agentic AI

For agentic AI systems that take autonomous actions in the real world, training data quality has amplified importance. An agent that has learned from biased, outdated, or poisoned data may make decisions that affect real users, real finances, and real systems. The security implications of training data quality scale with the autonomy and reach of the systems built on top of it.