Datasets:
language:
- en
task_categories:
- question-answering
tags:
- procedural-reasoning
- rule-reasoning
- retrieval
- synthetic-data
- large-language-models
pretty_name: RuleWorld
size_categories:
- 1M<n<10M
RuleWorld
RuleWorld is a large-scale benchmark for evaluating whether language models can retrieve and apply a shared repository of externally provided procedural rules. Its rules are abstract, globally reusable, and intentionally non-commonsense, so a model cannot answer reliably from world knowledge alone. Each rule is supplied in aligned natural-language (NL) and first-order logic (FOL) forms.
The benchmark covers three reasoning settings: Single-Rule QA, Parallel Multi-Rule QA, and Multi-Hop QA. Parallel questions combine independent sub-questions, while multi-hop questions require sequential rule application.
 |
 |
Dataset summary
| Resource | Records | Description |
|---|---|---|
| Rule library | 4,936,113 | Globally shared procedural rules, each in NL and FOL forms. |
| Full QA collection | 3,377,737 | 50,000 Single-Rule, 3,250,000 Parallel Multi-Rule, and 77,737 Multi-Hop instances. |
| Training split | 111,200 | 25,000 Single-Rule, 49,200 Parallel Multi-Rule, and 37,000 Multi-Hop instances. |
| Development split | 3,390 | Development instances. |
| Official evaluation set | 550 | The evaluation split: 550 instances. |
The full QA collection contains eleven fine-grained subtasks: single-rule, multi-rule-2 through multi-rule-8, and multi-hop-2 through multi-hop-4. For Multi-Rule QA, the numeric suffix denotes the number of distinct gold rules (with a minimum label of 2); for Multi-Hop QA, it denotes the reasoning depth.
Repository layout
all_data/
rule_library.txt # Full rule library; recommended rule source
rule_library.pkl # Serialized generation-time representation
single_qa.jsonl # 50,000 Single-Rule QA instances
multi_qa_base.jsonl # 30,000 base Multi-Rule QA instances
multi_qa_generated_{2,3,4}.jsonl # 3,220,000 composed Parallel Multi-Rule instances
multi_hop_qa.jsonl # 77,737 Multi-Hop QA instances
sampled/
train_111200.jsonl # Training split
dev_3390.jsonl # Development split
verified/
test_550.jsonl # Official balanced evaluation set
all_data/rule_library.txt is tab-separated with the columns rule_id, natural_language_rule, fol_rule, and rule_type. Use this text file rather than the pickle file when portability and security matter.
QA format
Every QA instance is a JSON object. Core fields are:
| Field | Description |
|---|---|
id |
Unique QA identifier. |
Q |
Natural-language question. |
A |
Gold answer list, ordered to match the question or its sub-questions. |
Q_template |
Prompt template containing the {rules} and {question} placeholders. |
A_template |
Reference answer template. After formatting, the answer is enclosed in \boxed{...}. |
proofs |
Gold derivation steps. Integer entries are rule IDs from the rule library. |
rule_info |
Structured gold rule metadata used for retrieval supervision. |
task_type |
One of single_rule, multi_rule, or multi_hop. |
qa_type |
Underlying rule subtype(s) involved in the instance. |
multi_rule records additionally contain rule_width, and multi_hop records contain rule_depth. Composed Multi-Rule records include source_ids, which preserve the order of their component QA instances. The official evaluation set also stores final_qa_type, selection_seed, and subtask.
Citation
If you find this dataset helpful in your research, we would kindly appreciate a citation:
@misc{yu2026factualknowledgebenchmarkinglearning,
title={Beyond Factual Knowledge: Benchmarking and Learning Step-Level Procedural Rule Reasoning in Large Language Models},
author={Bohan Yu and Pengfei Cao and Chen Han and Chenxi Zhou and Zhiheng Zhang and Zhiyang Xie and Wenhao Teng and Xiangwen Liao and Jun Zhao and Kang Liu},
year={2026},
eprint={2608.22753},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2608.22753},
}