SynthLabsAI/PRISM-Filter

Dataset Card for PERSONAS (Synthetic-Personas)

This dataset is a restricted subset of PRISM, focusing primarily on a questions where different personas could possibly have different answers. This dataset is to be treated as a research artifact, and should be used in tandem with the rest of the PERSONA dataset. The full PERSONA dataset is available here: https://huggingface.co/datasets/SynthLabsAI/PERSONA

Dataset Details

Dataset Description

The construction process for this dataset is outlined in the PERSONA (Prism Filter) dataset, but we'll copy it here for brevity.

Dataset Sources

As mentioned, the dataset is synthetically constructed. All prompts used can be found in the technical report. We also have a working benchmark that leverages the results of the paper to perform a personification evaluation.

• Paper: https://www.synthlabs.ai/research/persona

Uses

This dataset is intended to be used for personification research and data analysis into a wide spectrum of synthetic persona annotated preferences. We showed in our paper that there is a strong correlation between the dataset we present here and real human preferences, and as such we hope that this dataset can act as a surrogate into a number of more controversial human preferences.

Dataset Structure

This dataset follows a similar structure to the PRISM dataset -- every row is a different question found in the prism dataset. Answers are removed as they were not relevant to our study.

Dataset Creation

Please review our technical report for a thorough description of the dataset's construction.