sbordt/

TOAA-Membership-Inference

A paired member / non-member benchmark for membership-inference (MIA) experiments on the OLMo-2-179M models from Train Once, Answer All (TOAA). Each row is one sequence; members were inserted into the training data of the experiment model, non-members were not. Members and non-members are provided in an identical format so you can score both under a target and a reference model and run a test directly.

Target model (trained on the members):… See the full description on the dataset page: https://huggingface.co/datasets/sbordt/TOAA-Membership-Inference.