|
|
--- |
|
|
license: mit |
|
|
task_categories: |
|
|
- question-answering |
|
|
language: |
|
|
- en |
|
|
tags: |
|
|
- sports |
|
|
- travel |
|
|
configs: |
|
|
- config_name: corpus |
|
|
data_files: |
|
|
- "datasets/AI21-Hotels_eval/corpus.jsonl" |
|
|
- "datasets/AI21-Hotels_train/corpus.jsonl" |
|
|
- "datasets/AI21-WorldCup/corpus.jsonl" |
|
|
|
|
|
- config_name: questions |
|
|
data_files: |
|
|
- "datasets/AI21-Hotels_eval/questions.jsonl" |
|
|
- "datasets/AI21-Hotels_train/questions.jsonl" |
|
|
- "datasets/AI21-WorldCup/questions.jsonl" |
|
|
--- |
|
|
|
|
|
# AI21-Hotels and AI21-WorldCup datasets |
|
|
|
|
|
The **AI21-Hotels** and **AI21-WorldCup** datasets were created to support research on *aggregative question answering* in open-book settings. |
|
|
Aggregative questions require retrieving information from a large set of documents and applying reasoning over the collected text snippets. |
|
|
For example, the question “What is the fewest number of total goals scored in any single World Cup?” cannot usually be answered by a single passage. |
|
|
Instead, one must gather the goals scored in each tournament and then reason to identify the minimum. |
|
|
These datasets were first introduced in the paper "[Structured RAG for Answering Aggregative Questions](https://arxiv.org/abs/2511.08505v1)", which describes their design and construction in detail. |
|
|
|
|
|
## 🏨 AI21-Hotels dataset |
|
|
|
|
|
The **AI21-Hotels** dataset was automatically generated, with booking-like hotel pages, paired with aggregative queries. |
|
|
The train set (in folder AI21-Hotels-train) includes 50 documents, and 138 queries. |
|
|
The evaluation set (in folder AI21-Hotels-eval) includes 350 documents, and 193 queries. |
|
|
An example document from AI21-Hotels train set is shown in the image below: |
|
|
|
|
|
<p style="margin:0; padding:0;"> |
|
|
<img src="https://huggingface.co/datasets/ai21labs/aggregative_questions/resolve/main/additional_resources/hotels_train_doc.jpg" alt="Hotels Document" width=800/> |
|
|
</p> |
|
|
|
|
|
## ⚽ AI21-WorldCup dataset |
|
|
The **AI21-WorldCup** dataset consists of 22 Wikipedia pages, each corresponding to one WorldCup tournament held between 1930 and 2022, |
|
|
and 88 automatically generated aggregative queries. |
|
|
To increase the difficulty, the main statistics table was removed from each document. |
|
|
Because there are only 22 tournaments in total, the dataset is provided as a single collection without separate training and evaluation splits. |
|
|
|
|
|
|
|
|
|
|
|
## 📊 Statistics |
|
|
| Dataset | Split | # Docs | # Queries | Avg. Tokens / Doc | |
|
|
| -------- |-------| ------ | --------- |-------------------| |
|
|
| Hotels | Train | 50 | 138 | 592 | |
|
|
| Hotels | Eval | 350 | 193 | 596 | |
|
|
| WorldCup | --- | 22 | 88 | 18881 | |
|
|
|
|
|
|
|
|
## Languages |
|
|
English |
|
|
|
|
|
|
|
|
## License |
|
|
Datasets are released under the MIT license. |
|
|
|
|
|
## Citation |
|
|
If you use these datasets, please cite: |
|
|
```bibtex |
|
|
@article{koshorek2025structured, |
|
|
title={Structured RAG for Answering Aggregative Questions}, |
|
|
author={Koshorek, Omri and Granot, Niv and Alloni, Aviv and Admati, Shahar and Hendel, Roee and Weiss, Ido and Arazi, Alan and Cohen, Shay-Nitzan and Belinkov, Yonatan}, |
|
|
journal={arXiv preprint arXiv:2511.08505}, |
|
|
year={2025} |
|
|
} |
