use case · wild
A model learns from human text, and human text carries human associations. Here you write one
template with a {X} you swap and a [MASK] you predict, then run it across
several subjects at once. Line the predictions up side by side and the training-data associations
become visible — sometimes uncomfortably so. This is the same method researchers use to audit
model bias, running live on your device.
This isn't the model "being offensive" — it's a mirror of statistical patterns in the text it was trained on. Seeing it plainly is the point: you can't mitigate a bias you can't measure, and doing the measurement locally means you can probe sensitive templates without sending anything to a server.
← Practical · Model · Next: Multi-model →