Start with the task, not the data
Before you collect anything, write down what the model should do, what a correct answer looks like, and where it will be used. “Detect damaged products in return photos” leads to very different data than “answer customer questions about our products”. The task decides which data is useful and which is noise.
1. Collect from sources you are allowed to use
List every source: your own systems, customers, partners, public datasets, scraped web pages. For each, check that you have the right to use it for training, and for personal data, that you have a legal basis under the GDPR. Keep the list; you will need it later.
2. Clean what you collected
Check every file for broken files, duplicates, low quality and personal data, and record what you find. For images that means blur, exposure and copies; for video frozen and dark footage; for text personal data, exposed keys and copies. See how to clean an AI training dataset.
3. Label consistently
If the model learns from labels, write labelling guidelines with examples of hard cases, and measure how often two labellers agree. Inconsistent labels limit the model no matter how much data you add. See label quality.
4. Split before you train
Set aside validation and test data before training, and split by source (customer, camera, document) rather than by file, so near copies do not end up on both sides. A test set that leaks into training gives scores you cannot trust.
5. Start small and look at the errors
Train a first version on a modest, clean subset, then read the model’s mistakes. They tell you which data to add next far better than a bigger, messier collection would.
6. Document the dataset
Write down sources, cleaning steps, labelling rules, splits and known gaps. The record lets you reproduce the model and answer questions from customers and regulators. See dataset documentation.