The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.

Deriva test set: nothing is released yet

This repository holds a card and no data files.

Content warning. The study behind this repository is about harmful requests. This repository contains none of them.

Status (2026-11-16)

The five-language safety test set of the Deriva study (harmful prompts and harmless-but-edgy prompts in English, Hindi, Hinglish, Arabic and Swahili) is not frozen: its translation into the four other languages did not finish. Nothing is released before the freeze. No model was run, so there are also no judged replies, labels, tables or figures.

What this dataset is planned to hold

  • the prompts, each row with its source and licence;
  • one label per judged reply, with an excerpt cut to 200 characters;
  • one correct or incorrect mark per task question;
  • the tables and figures.

Not to be released: any model reply beyond the 200-character excerpt, and any item of an evaluation-only source.

Sources and licences

Every source, with its licence, is listed in the article under "Datasets, benchmarks and licenses" and in data/sources.yaml of the code repository.

Links

Downloads last month
24

Collection including Jemiiah/deriva-testset