This dataset contains results from all rounds of Adversarial Nibbler. This data includes adversarial prompts fed into public generative text2image models and validations for unsafe images. There will be two sets of data: all prompts submitted and all prompts attempted (sent to t2i models but not submitted as unsafe).

WIT (Wikipedia-based Image Text) Dataset is a large multimodal multilingual dataset comprising 37M+ image-text sets with 11M+ unique images across 100+ languages.

This dataset contains synthetic training data for grammatical error correction. The corpus is generated by corrupting clean sentences from C4 using a tagged corruption model. The approach and the dataset are described in more detail by Stahlberg and Kumar (2021) (https://www.aclweb.org/anthology/2021.bea-1.4/)

SeeGULL Multilingual is a multilingual and multicultural dataset of stereotypes. It consists of stereotypes in 20 languages with human annotations across 23 languages, including annotations on their degree of offensiveness.

ToTTo is an open-domain English table-to-text dataset with over 120,000 training examples that proposes a controlled generation task: given a Wikipedia table and a set of highlighted table cells, produce a one-sentence description. We hope it can serve as a useful research benchmark for high-precision conditional text generation.

IndicGenBench is a high-quality, multilingual, multi-way parallel benchmark for evaluating Large Language Models (LLMs) on 4 user-facing generation tasks across a diverse set 29 of Indic languages covering 13 scripts and 4 language families.

The HierText dataset contains ~12k images from the Open Images dataset v6 with large amount of text entities. We provide word, line and paragraph level annotations.

The CF-TriviaQA dataset accompanies "Hallucination Augmented Recitations for Language Models" paper (https://arxiv.org/abs/2311.07424). It is a counterfactual open book QA dataset generated from the TriviaQA dataset using Hallucination Augmented Recitations (HAR) approach, with the purpose of improving attribution in LLMs.

The BamTwoogle dataset accompanies "ReST meets ReAct: Self-Improvement for Multi-Step Reasoning LLM Agent" paper (https://arxiv.org/abs/2312.10003). It was written to be a complementary, slightly more challenging sequel to Bamboogle dataset. It addresses some of the shortcomings of Bamboogle we discovered while performing human evals for the paper.