Datasets:
dataset_info:
- config_name: corpus
features:
- name: id
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 56944692
num_examples: 300000
download_size: 31526767
dataset_size: 56944692
- config_name: counting
features:
- name: id
dtype: string
- name: text
dtype: string
- name: atoms
list: string
splits:
- name: train
num_bytes: 5647151
num_examples: 13020
download_size: 4695849
dataset_size: 5647151
- config_name: queries
features:
- name: id
dtype: string
- name: atom
dtype: string
- name: sentence_id
dtype: string
splits:
- name: train
num_bytes: 3952670
num_examples: 36337
download_size: 2137129
dataset_size: 3952670
- config_name: source_sentences
features:
- name: id
dtype: string
- name: text
dtype: string
- name: n_atoms
dtype: int64
splits:
- name: train
num_bytes: 2886104
num_examples: 14483
download_size: 1749155
dataset_size: 2886104
configs:
- config_name: corpus
data_files:
- split: train
path: corpus/train-*
- config_name: counting
data_files:
- split: train
path: counting/train-*
- config_name: queries
data_files:
- split: train
path: queries/train-*
- config_name: source_sentences
data_files:
- split: train
path: source_sentences/train-*
license: cc-by-sa-4.0
task_categories:
- text-retrieval
language:
- en
pretty_name: PropSearch
size_categories:
- 100K<n<1M
PropSearch
This benchmark measures the ability of proposition-level embedding models (as well as fixed-size sentence encoders) to retrieve a source sentence starting from an extracted atomic proposition. The benchmark is built on sentences extracted from scientific abstracts (arXiv abstracts and papers-with-abstracts).
Atomic propositions are extracted by Claude Sonnet 4.6 using the following system prompt:
Extract atomic propositions from the given sentence. The propositions need to be independent (for instance, no pronouns that refer to entities outside the proposition itself) and connectives between propositions need to be discarded.
If the sentence only has one atomic proposition, return a singleton.
Return the output as a Python list.
In total, the task has 14,483 source sentences, 36,337 query propositions and a corpus of 300,000 confounding sentences.
For the full data creation and validation procedure, check our EMNLP 2026 paper Variable-size Discovery and Representation of Atomic Propositions.
We release this data for evaluation purposes. In the creation of this benchmark, we use Claude Sonnet 4.6 for data generation. Therefore, as to Anthropic's Privacy and Legal Policies, this data must not be used to train models that are competitive with Anthropic's own.
Subsets
The dataset contains four subsets:
source_sentences: The source sentences, along with an identifier and their original number of atomic propositions (n_atoms, used to stratify results)queries: The extracted propositions, each labelled with the correct source sentence IDcorpus: A large number of confounder sentences, with no corresponding atomic propositionscounting: Sentences with their full list of atomic propositions, used to evaluate proposition counting
Data Distribution
Source Sentences by Atom Count
| Atom Count | Total Sentences |
|---|---|
| 1 | 2,089 |
| 2 | 5,233 |
| 3 | 3,516 |
| 4 | 2,116 |
| 5 | 961 |
| 6+ | 952 |
Query Atoms by Source-sentence Atom Count
| Atom Count | Total Query Atoms |
|---|---|
| 1 | 1,962 |
| 2 | 9,195 |
| 3 | 8,850 |
| 4 | 7,040 |
| 5 | 3,921 |
| 6+ | 5,369 |
Citation
If you use this task, please cite us as:
@inproceedings{ortame-navigli-2026-variable,
title = "Variable-size Discovery and Representation of Atomic Propositions",
author = "Ortame, Francesco and
Navigli, Roberto",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.2302/",
pages = "52330--52348",
ISBN = "979-8-89176-459-0"
}