Our two goals with Rampart Edge were 1. Reduce false positives 2. Make it way smaller We built this model to prevent PII from ever even leaving a user's device on America.gov. Meaning it needs to be downloaded on every device before chat messages leave To do this we began with a 50.74MB base model and hill-climbed with the following big wins: - Reducing from 6 layers to 4 layesr - Making the word table smaller - 4-bit weights, shared between layers The training data was entirely synthetic. We saw big wins by starting with government redacted documents and filling in synthetic PII.

Announcing American Forge. A six-month residency for hardcore engineers to join the US government. Created in partnership @ndstudio x @cognition

My first task for the National Design Studio was to make Rampart smaller. It's our frontier on-device privacy model trained to filter all personal user data before it even gets to America.gov After a lot of hill-climbing and data labeling, I'm excited to launch Rampart Edge along with @as400495. So small the weights fit in a notebook. This will noticeably speed up the load time of America.gov for hundreds of thousands of Americans. Under budget and ahead of schedule!