Data labelling is the work of people preparing the examples that AI models learn from, by attaching the correct answer, or label, to each piece of data. For language models, it also includes writing model answers and giving human feedback, in which an expert reads two answers from a model and marks which is better. It is also called data annotation. Labelling companies hire large numbers of trained people, often experts in a field, and pay them per task or per hour.
A model learns a task by seeing many examples of it done correctly. For a speech model, an example is a recording paired with an exact transcript. For an image model, it is a photo with each object outlined and named. For a large language model (LLM), two kinds of example matter most after the first stage of training. The first kind is a question written together with a correct, complete answer, which teaches the model to follow instructions. The second kind is a ranking. The model writes two answers, and an expert says which is better and why. The model is then trained to give more answers like the preferred ones. This is called reinforcement learning from human feedback, or RLHF.
Quality is the whole job, because a wrong label teaches the model a wrong answer. So labelling companies test each annotator before hiring. They give two people the same task and compare their labels, and they send any disagreement to a senior reviewer. For specialist subjects, they hire specialists: doctors for medical questions, lawyers for legal ones and native speakers for each language.
Here is an example. A model maker wants its model to answer health questions better in Hindi. A doctor in Lucknow, who works on a labelling platform in the evenings, receives a question about when a dengue patient should go to hospital. The model has written two answers. The doctor marks the second answer as better, corrects one wrong number in it and writes two lines explaining why. The task takes eight minutes, and she is paid for each task. Thousands of such judgements, from hundreds of doctors, go into one round of training.
Indian labelling teams do this work in many languages. For speech and voice AI, they transcribe recorded speech word by word, marking each speaker and each pause. The finished data is used by training jobs, which often run on a distributed computing framework. The same experts write the approved answers in the test sets used for LLM evaluation.
How is labelling different from intelligent document processing (IDP)? IDP uses a trained model to read documents. Labelling is the human work that creates the examples such models learn from.