timo-pierre-schrader/canonical-order-problem
List Extraction Dataset A dataset for studying how large language models (LLMs) represent and generate multi-valued relations — relations in which a single subject is associated with a set of entities (e.g., "all countries in South America", "all elements of the periodic table"). It accompanies the paper: The Canonical Order Problem: When Large Language Models Are Unreliable Knowledge Bases for Multi-Valued Relations Timo Pierre Schrader, Annemarie Friedrich, Simon Razniewski… See the full description on the dataset page: https://huggingface.co/datasets/timo-pierre-schrader/canonical-order-problem.
List Extraction Dataset
A dataset for studying how large language models (LLMs) represent and generate multi-valued relations — relations in which a single subject is associated with a set of entities (e.g., "all countries in South America", "all elements of the periodic table").
It accompanies the paper:
The Canonical Order Problem: When Large Language Models Are Unreliable Knowledge Bases for Multi-Valued Relations Timo Pierre Schrader, Annemarie Friedrich, Simon Razniewski, Lukas Lange
The Canonical Order Problem
LLMs are increasingly used as knowledge bases. While much prior work focuses on extracting single relational triples, most real-world relations are multi-valued and require generating a complete set of entities.
This dataset is built to investigate the canonical order problem: the probabilistic distributions inside LLMs organize many multi-valued relations according to a canonical ordering (e.g., alphabetical, chronological, numerical, or another domain-specific convention). As a result, prompts that deviate from this internal canonical ordering lead to a markedly reduced reliability of LLMs when generating complete sets.
Each example provides a prefix (a subset of the target set) and asks the model to complete the remaining entities. By varying how the prefix is ordered, the dataset isolates the effect of deviating from the canonical order on set-completion reliability.
Topics
The dataset covers 37 distinct multi-valued relations (topics), each with a well-defined, stable number of entities, plus 1 extreme topic used for stress-testing. Topics are categorized by set size:
Each topic is assigned a canonical ordering — the most common convention in authoritative sources (e.g., Wikipedia, official lists): alphabetical, chronological, numerical, or other.
Topics span a wide range of domains, e.g. Nobel prize categories, planets of the solar system, German states, NBA teams, U.S. presidents, books of the Bible, constellations, elements of the periodic table, and (for the extreme case) prime numbers up to 100,000.
Dataset Structure
The dataset is organized into 5 configurations (one per ordering strategy), each with train, validation, and test splits.
Note: Thetrain,validation, andtestsplits contain identical data. They are provided for API compatibility; any single split can be used for evaluation.
Configurations
The configuration name is english_<ordering>, e.g. english_shuffled.
For ordered, one example is generated for each prefix length k = 1 … (n−1). For one_after_another, one example is generated per entity. For shuffled and non_shuffled, a fixed set of random subsets of varying lengths is used per topic.
Row counts (per split)
Features
The generation task is: given prefix_entities (or its formatted variants), generate expected_entities to complete the full set list_entities.
Citation
If you use this dataset in your research, please cite the paper:
TBALicense
This dataset is released under the CC BY-SA 4.0.
