---
license: mit
---
This dataset is a deduplicated subset of the train split of [PiQA](https://huggingface.co/datasets/ybisk/piqa), as used in the paper [How Much Can We Forget about Data Contamination?](https://arxiv.org/abs/2410.03249). The deduplication was performed using [this script](https://github.com/tml-tuebingen/forgetting-contamination/blob/main/evaluation/find_duplicate_benchmark_questions.py).

The data fields are the same as in https://huggingface.co/datasets/ybisk/piqa, with the additional "split-id" column that can be used to partition the benchmark questions into different subsets.
The dataset can be used as a plug-in replacement if you want to work with the deduplicated benchmark questions.

For details on how the dataset was constructed, see Section A.4 and Table 2 in [the paper](https://arxiv.org/abs/2410.03249).