Announcement
Dataset watermark aims to reveal whether an AI model trained on marked data
The researchers say it can be detected from model-generated text with API-only access and offers a provable false-positive guarantee.
TLDR
The researchers introducing the watermark say it could help determine whether a model trained on a marked benchmark or dataset using only text the model generates through an API. They say it preserves the data’s utility and semantics and offers a provable false-positive guarantee.
Combined views
250
2 Sources, first seen ago
