Label

A person attaches the label

A label isn't a value already sitting inside the material. A photo file contains only pixels — the word "cat" doesn't appear anywhere in it. A person looked at the photo, made a judgment, and wrote it down alongside.

That means making labels is work that eats a person's time directly. Naming ten thousand photos means looking at ten thousand photos and writing ten thousand words down. A task that also has to mark where an object sits can take several minutes per photo.

What gets chosen as the answer depends on the goal. The same photo could get labeled "cat," or "indoors," or "bright." A label isn't a property of the material — it carries the problem being solved.

Grading is only possible once there's a label

Training is grading, done over and over. The model looks at the material and gives an answer; that answer gets checked against the label sitting next to it. If it's wrong, the size of the error gets measured, and the numbers inside the model get nudged by that much. Repeat that countless times and skill builds up.

Without a label, none of that grading can happen. There's no way to know whether the answer was right or wrong, so there's no way to know which direction to correct. That's exactly why training on labeled material is called supervised learning.

Features and labels sit in different spots

Picture one record as a row in a table: the earlier columns are features, and the last column is the label. A feature is what the model looks at; a label is what the model has to guess correctly. The two have to be paired up before any of it can go into training.

Once a model is actually being used, though, there's no label. A new photo arrives with only features attached, and figuring out the answer is the model's job. Labels are only needed while teaching and while measuring skill.

That's why a value that can only be filled in once you already know the answer must never end up mixed into the features. Test scores look great, but the model turns out useless in practice, because that value can't actually be had out in the real world.

Get the label wrong and the training goes wrong

A model never questions the label attached to it. If someone accidentally writes "cat" on a photo of a dog, the model learns it exactly that way. Even a small percentage of wrong answers mixed in can drag performance down noticeably, and it's hard to trace what caused it.

Plenty of cases are wrong simply because they're genuinely ambiguous. Something like "is this review positive or negative" gets judged differently from person to person. That's why teams write the standard down as a document beforehand and have several people label the same material to check how well their judgments line up. A low agreement rate is a sign the standard needs fixing, not the material.

There's also a path that learns without labels

Since attaching labels is so expensive, several ways to cut that cost have emerged. One groups unlabeled material by similarity to itself. Another combines a small pile of labeled material with a much larger pile that isn't labeled at all.

The approach today's language models lean on is close to this too. Instead of a person writing down the answer, part of a passage gets hidden and the model has to guess what belongs there. The text that was already there does the job of an answer, so no human hours are spent. Even then, the final stage of polishing a model to match human taste still needs a person's judgment attached.