What “training a model on your data” actually means

You don't upload a CSV and press train. First we clean the data, then we train, then we set the thresholds.

Training a model on your data is not uploading a CSV and pressing “train”. We start by cleaning up the sources: removing duplicates, merging channels, and checking that the labels really reflect the decisions your moderators make.

Then we build a test set from real traffic: short comments, longer posts, borderline content, and cases where something went wrong before. Without it, the numbers look good on paper and the model stumbles in production.

Training is only half the work. The other half is thresholds: where ALLOW ends and REVIEW and BLOCK begin, and what those thresholds do to your team's workload. We look for a sensible compromise, not the best score on a single metric.

After launch, cases from the REVIEW queue and the mistakes (false alarms and misses) go back into further corrections of the model. The model then understands the language and the abuse patterns in your company better.

A model that stops learning after launch goes stale quickly. That is why corrections are part of the service.