Fine-tuning labs
Train on data that behaves like the original, but isn't.
Fine-tuning labs need data that behaves like the real thing, with no real person left in it. Models trained on data that keeps its meaning are more accurate, and more useful to the people who rely on them.
Both sides of the sale
For the provider and the lab that buys from them
We work with both the data provider and the labs buying the data, so the dataset that changes hands is safe to sell and still worth training on.
Nakato reads each record for meaning. It finds every direct identifier and every indirect one, the contextual details that point to one person when put together, and gives each a twin of its own. Every twin is contextually equivalent, so a model can be trained on it as if it were real. See our benchmarks.
Record · Support conversation
| Field | Original stays with you | Twin what the model sees | Preserved |
|---|---|---|---|
| Customer | Priya Shah | Anika Rao | gender, name origin |
| Account | 40912345 | 71520984 | valid format |
| Home town | Leeds | Sheffield | region, city size |
| Issue | Card declined abroad | Card declined abroad | unchanged |
| Resolution | Travel flag added, fee refunded | Travel flag added, fee refunded | unchanged |
Why twin
Data is valuable because it is about people
Proprietary data is valuable precisely because it is proprietary: it is bound up with the people it describes. Strip the people out and you strip the value with them. Masking, tokenisation and pseudonymisation.
So we built semantic twinning to keep the individuality in data. A twin keeps a record's context, so it keeps its usefulness, and changes the person it is about. It also gives a model something to reason with: data that keeps its shape but changes its face gives markedly better results in downstream tasks than masked data. Our results.
Real enough to give real results. Changed enough that no one can see the people beneath it.
The twin keeps what a model learns from and changes who it is about.