Skip to main content
Rows from real work, such as support tickets and chat logs, can hold names, addresses, and account numbers. monte dataset add --anonymize replaces those details before the rows are stored, so only anonymized rows reach artifact storage. Anonymization removes personal details. It does not grant the right to train on the rows.

​
Anonymize Rows

--preview prints rows before and after, and stores nothing. Run the command again without --preview to add the dataset. Two detectors find the details. A named entity recognition (NER) model finds names in context, and pattern rules find structured identifiers. The NER model runs where monte dataset add runs, so no row goes to a third-party model. --entities limits the types, and --threshold (default 0.5) sets the lowest NER confidence that counts as a find. A lower threshold finds more names and also replaces more words that are not personal.

​
Choose a Replacement

A placeholder or a pseudonym maps the same value to the same replacement in every turn of a row, so a ticket that names the customer twice still reads as one person:

​
Keep Terms and Add Patterns

A keep file lists names that must stay, such as the company’s own product names. A pattern file adds identifier types that the built-in rules do not know. Each key becomes a type with its own placeholder.
keep.txt
patterns.yaml
With these files, CASE-104233 becomes [CASE_1], and Acme Cloud stays as it is.

​
Check What Remains

monte dataset show support-replies prints how many values of each type the add replaced.
Detection misses some details, and a model can learn and repeat a detail that stayed in its rows. Read a sample of the rows before a run trains on them. A fix is a new version of the dataset.

​
Storage

With BYOC, the rows stay in the team’s artifact storage, and generation and training run on GPU instances in BYOC, so no row passes through Monte. With Monte Cloud, Monte stores the rows and its GPU instances read them. Security and Privacy lists what Monte can see in each option.

​
Delete a Dataset

Deleting a dataset does not remove what a model learned from it. To remove that influence, delete the checkpoints trained on it and their deployments, then train again without the dataset.
The command removes every version from artifact storage. Run history keeps the record that a run trained on the dataset, but not the rows.