The work before training: cleaning, structuring and checking datasets so a model has something honest to learn from.
This is the least visible of the three services and the one that most often decides whether a model works. It covers everything between the data you have and a dataset a model can be trained on honestly: cleaning, structuring, splitting, checking, and documenting what was done.
The reason it matters is that broken datasets do not fail loudly. They produce evaluation numbers that look fine and a model that disappoints in production, which is the most expensive way to find out.
Four kinds of task, usually in this order, though an audit often sends us back a step.
The recurring faults. Every one of these produces a model that scores better in evaluation than it behaves in production, which is why they are worth finding before training rather than after.
A dataset audit is usually the cheapest useful thing we can do for a project, and it can be done before any labelling is commissioned.
What exists, in what formats, with what labels, and how much of it. Frequently the first useful output, because the answer is often not what the project assumed.
Duplicates, near-duplicates, leakage across splits, encoding damage, records that are technically valid and semantically empty.
A sample re-labelled blind against the guideline and compared, so label noise gets a number instead of an opinion.
Written up with severity and what each fault would cost if left alone. You decide what to fix; we will say which items we think matter most.
The agreed fixes, applied with a record of every change rather than a silently overwritten dataset.
The cleaned dataset with its splits, the guideline it was checked against, and a written record of what changed and why. The record is the part people underestimate: without it, the next person cannot tell a deliberate decision from an accident, and the dataset slowly becomes untrustworthy.
Where we find something that cannot be fixed by cleaning, such as a class that simply has too few real examples, that goes in the findings as a limitation rather than being papered over with augmentation.
Processing inside Swiss borders is available on request rather than by default, because it depends on which tools a project needs and where the people doing the work are. Ask when you scope the project and we will confirm in writing what is possible for your data before anything moves.
If a project has no residency constraint, saying so early usually widens what we can offer and shortens the setup.
Yes, and it is often the sensible first step. The audit tells you whether the dataset can support the model you want. Some projects take the findings and fix things internally, which is a good outcome.
By similarity rather than exact matching, because exact-match deduplication misses the cases that matter: the same document reformatted, the same photograph resized, the same record entered twice with a different identifier. Those are precisely the ones that leak across a split.
Sometimes for invariances you can describe, such as rotation or lighting. Rarely for a class that is genuinely rare, because augmenting five examples produces variations of five examples rather than knowledge of the class. We will say which case you are in.
It depends on the data and the regime, and it is a scoping conversation rather than a yes or no on a web page. Bring the constraint early: it changes who can touch the data, which tools are usable, and where processing can happen.
Most projects touch more than one of these. One partner across all of them means the labelling, the preparation and the application are not three vendors blaming each other.
Human-labelled training data across text, image, video, audio and 3D, and the feedback layer for language models: preference data, written demonstrations, rubric scoring, red-teaming.
Custom AI-driven applications, taken from concept through to deployment by the team that labelled the data.
Rendered training data built in 3D instead of collected and labelled: for classes that are rare, expensive to photograph, or cannot leave the building.
