ongsici/Dolma-Paloma
Dataset Card for Dolma-Paloma Dataset Summary This dataset was created as part of my Master's thesis research on "Leveraging Model Checkpoints for Membership Inference Attacks on Large Language Models". The aim was to create a clean setup, free from distribution shifts, to study proposed checkpoint MIA methods and their performance using checkpoints from Pythia models. It was derived from the original Dolma and Paloma datasets by applying reservoir sampling and… See the full description on the dataset page: https://huggingface.co/datasets/ongsici/Dolma-Paloma.
Dataset Card for Dolma-Paloma
Dataset Summary
This dataset was created as part of my Master's thesis research on "Leveraging Model Checkpoints for Membership Inference Attacks on Large Language Models". The aim was to create a clean setup, free from distribution shifts, to study proposed checkpoint MIA methods and their performance using checkpoints from Pythia models. It was derived from the original Dolma and Paloma datasets by applying reservoir sampling and data processing steps.
Data Processing
Random sampling was applied to Paloma (Dolma v1.5 domain) dataset to obtain domains with >100 non-members from val and test splits that with sequence lengths >=2048. Reservoir sampling was applied to Dolma v1.5 to obtain >100 members for the same domains from train split with sequence lengths >=2048.
Samples were tokenised using Llama-2-7B model and samples with above 2048 token lengths were retained. Sample tokens were then truncated to various sequence lengths: [64, 128, 256, 512, 1024, 2048], and the sample sequences were then regenerated from the tokens.
This was applied to 3 domains: Wikipedia, Stack, and Pes2o.
Dataset Information
Intended Use
The dataset is intended for academic research purposes, particularly for evaluating methods for MIAs.
