---
task_categories:
- question-answering
- zero-shot-classification
language:
- en
size_categories:
- n<1K
---

### Dataset Description

This dataset contains binary-choice questions about shared materials between objects (mostly artifacts). To correctly answer the questions, one needs to consider detailed information about the possible materials that the objects can be made of. To construct our dataset, we began with an extensive schematic dataset detailing the parts and material composition of primarily man-made objects. From this dataset, we identified triples of objects where the first two objects share a common material, while the third object does not share any material with the second object. We then manually selected 100 triples from this set and corrected any inaccuracies to generate a curated set of test questions for our new dataset.

Our questions ask which object is less likely to share material with the second object. This phrasing is preferable to asking which artifact is more likely to share materials with the target object, as it avoids the connotation that the amount of shared material is quantitatively significant for the correct answer.

The first two objects of a triple share common materials. When we formulate the questions, we ensure that the correct answer for the first 50 questions is 'b)', and 'a)' for the last 50 questions to keep the test fair and avoid bias.


## Uses

This dataset can be used to test large language models.


## Bias, Risks, and Limitations

The dataset is small, with only 100 items.

The common materials may not be exhaustive lists of any possible material shared between any instances of the first two objects of a triple. Not every pair of instances of the first two objects have a common material. The common materials are materials that the two objects may share. They provide references when one is not sure about the answer.

## Paper

This dataset is created for the experiment section of the following paper: Language Models Benefit from Preparation with Elicited Knowledge