Data Verification and Quality Checks

A model learns whatever patterns are present in the examples it receives. If those examples contain stray markup, content the model should never use, or records that are simply poor, the model absorbs those problems along with everything useful. This means the people preparing training data must decide, before training begins, what stays, what goes, and what might be worth keeping in a smaller form.

This article follows that decision-making from start to finish. It covers cleaning data that is not usable as it stands, detecting records that are low in quality, and filtering a collection that is larger than the project can afford to use.

Many public datasets are assembled by collecting text from across the internet. Web pages are written in markup languages, so text pulled from them often arrives wrapped in HTML tags. These tags may describe page structure, add styling hints, or mark links. Unless the goal is to teach a model to read those tags, they are noise.

The first cleaning task is therefore to remove extraneous formatting tokens, including HTML and Markdown markup that ended up inside the text itself. The source pack summarizes the benefit directly: removing these tokens can improve accuracy while also cutting the length of the input.

In the diagram, the tags on the left are marked as changed because they are the parts removed. Only the words survive on the right. The tags were never part of what the model should learn to produce, so their removal changes nothing useful about the text itself.

A simple way to strip angle-bracketed tags is a regular expression that matches anything between a less-than sign and a greater-than sign. The example below is an illustration of the idea, not a full cleaning pipeline.

Cleaning also means enforcing rules. Data has to be checked against the policies that govern the project, and anything that does not comply has to be removed. The source names several categories that fall under this heading: personally identifiable information (PII), sensitive data, copyrighted material, and content considered toxic.

Techniques for handling these categories are covered in an earlier chapter on privacy and safety. This article focuses on the filtering step itself: the point in the pipeline where non-compliant material is actually taken out.

After compliance cleaning, the data may still contain records that are poor in quality even though they are allowed and properly formatted. A record can be clean in form and still be a weak example for learning. Finding these records relies on a combination of two activities: looking closely at the data by hand, and turning what is noticed into rules that can be applied across the whole collection.

Manual inspection reveals patterns. Once a person notices that a certain kind of record is poor, that pattern can become a heuristic, a rule that flags similar records automatically. The source notes that some heuristics depend on context, such as the record's position within an annotation session. A rule like that cannot be judged from the record alone; it depends on where the record sits in the process that produced it.

The two activities feed each other. Manual review finds the pattern, the heuristic applies it at scale, and further review checks whether the rule is catching the right records. The diagram below shows the order of the cleaning stages that surround this step.

The steps described so far are not independent choices. They form a sequence in which each stage hands a smaller and cleaner collection to the next. Raw scraped data first has its formatting tokens removed, then is checked against project policies, then has non-compliant fields dropped, then is screened for low-quality records. Only after that, if the project has more data than it can use, does an optional filtering stage select a subset.

Read the diagram from left to right. Each stage removes something, so the collection on the right is smaller and cleaner than the one on the left. The final filtering stage is marked as a decision because it applies only when there is more data than the project can afford to use.

Sometimes the problem is not bad data but too much of it. If there is more data than you need or can afford to use, for example because of your compute budget, you can filter further. The source names two techniques for choosing which examples to keep.

Active learning works as a cycle rather than a single choice. The model is trained, a selection step picks the examples it would learn most from, a human reviewer labels those examples, and the model is retrained on the result. The cycle then repeats. The diagram below shows the order in which these parties interact across rounds.

Importance sampling is different in emphasis. It keeps the examples judged most important to the task, and it does not depend on repeated rounds of retraining in the way the active learning loop does. In both cases, the subset is chosen deliberately rather than by chance.

Both techniques depend on a single requirement, which the source states directly: their efficiency depends on whether you have a good way to evaluate the importance of each training example. Choosing examples is only as good as the measure used to judge them. A weak measure keeps the wrong examples; a strong measure lets a much smaller set carry most of the value.

Consider a simple cut-off made of a score for each record. Every record scoring above the threshold is kept and every record below it is dropped. If the score is accurate, the records above the line are the ones that matter, and a large share of the collection can be removed safely. If the score is noisy, good records can fall below the line and be lost, while poor records can climb above it and stay in the set.

The source ties the value of this work to cost. Meta researchers, in their study of data pruning (Sorscher and colleagues, 2022), concluded that discovering good data-pruning metrics can significantly reduce the resource costs of modern deep learning. The practical message is that a good metric is an investment, because it determines how much of a large dataset can be dropped without losing the examples that matter.