🧠 SLR-Bench-Dutch: Scalable Logical Reasoning Benchmark (Dutch Edition)

SLR-Bench Multilingual Versions:

SLR-Bench-Dutch is the Dutch-language pendant of the original SLR-Bench dataset. It follows the same symbolic structure, evaluation framework, and curriculum as the English version but provides all natural-language task prompts translated into Dutch.

This enables systematic evaluation and training of Large Language Models (LLMs) in logical reasoning in Dutch, supporting both multilingual reasoning and cross-lingual generalization research.

DS Overview

• Curriculum: 20 complexity levels, grouped into 4 broad tiers (basic, easy, medium, hard)
• Tasks: >19,000, each comprising: A natural language prompt, an executable validation program for automatic evaluation, and a latent ground-truth rule.
• Application: SLR-Bench can used to evaluate conventional and reasoning LLMs (e.g., GPT-4o, Llama-3, Gemini, DeepSeek-R1) and to train models via curriculum learning.

Key Features of SLR

• 🔨 Automatic Task Generation: Synthesize new inductive reasoning tasks with controllable complexity, novel logic rules, and natural language prompts—no need for human annotation.
• 🧩 Programmable & Scalable: Specify your own logic vocabulary, grammar, rule distributions, and task parameters; supports curriculum-style scaling and out-of-distribution task creation.
• 🧠 Symbolic, Automated Evaluation: Deterministically verify LLM outputs via the validation program, not MCQA, LLM judge, or exact matching.
• 📈 Curriculum Learning: Use SLR-Bench, a structured 20-level benchmark, for evaluating and training models across a span of logical challenges.

---

Quick Start

Loading the Dataset

from datasets import load_dataset
# Load SLR-Bench test split
ds = load_dataset("AIML-TUDA/SLR-Bench-Dutch", "v1-All", split="test")

Evaluate using SLR-Bench

Requires the evaluate library and a Prolog interpreter installed on your system (e.g., SWI-Prolog). Install the required dependencies via:

pip install evaluate
sudo apt-get install swi-prolog

Example Usage

from evaluate import load
symbolic_judge = load("AIML-TUDA/VerifiableRewardsForScalableLogicalReasoning")
rules = ds["ground-truth rule"]  # For demo only—use model predictions in practice
references = [
    {
        "validation_program": p,
        "evaluation_config": {
            "positive_predicate": "oost",
            "negative_predicate": "west"
        }
    } for p in ds["validation program"]
]

results = symbolic_judge.compute(predictions=rules, references=references)
print(results)

Note: For real evaluation, replace rules with your model's predicted rules. Here, we use ground-truth rules for demonstration only.

Example results:

{'accuracy': 1.0,
 'partial_score': 1.0,
 'syntax_score': 1.0,
 'detailed_results': [{'is_correct': True,'partial_score': 1.0,'syntax_valid': True,'error': None,'exec_time1': 0.014362812042236328},
                      {'is_correct': True,'partial_score': 1.0,'syntax_valid': True,'error': None,'exec_time1': 0.012364625930786133}]
}

---

Dataset Columns

Column Name	Type	Description
id	int64	Unique identifier for each dataset entry (row).
prompt	string	The instruction prompt of the logical reasoning task.
ground-truth rule	string	The latent logical rule that solves the given task.
validation program	string	The executable logic program used by the symbolic judge to verify candidate model solutions for the task.
symbols	string	Symbolic representation of the bckground knowledge
curriculum level	int64	The specific level (1-20) in the SLR-Bench curriculum that this task belongs to, reflecting difficulty.
curriculum tier	string	The broader difficulty tier grouping multiple levels (e.g., "basic", "easy", "medium", "hard").
rule sampling	string	The policy or method used to generate the ground-truth rule (e.g., "uniform", "llm-guided").
rule complexity	string	The length of the logic rule, counting the number of used predicates without the has_car predicate.
background sampling	string	The policy used to sample background knowledge for the task (e.g., "mirror", "uniform").
problem size	int64	Total number of labeled examples (positive + negative) provided in the task instance.
vocabulary predicates	int64	Number of unique predicate symbols available in the vocabulary for constructing rules and background knowledge.
vocabulary car constants	string	List of car constant symbols (e.g., "car1", "car2", ...) available in the vocabulary for the task.

---

SLR-Bench Curriculum

Stage	Level	#Consts	#Preds	κ (Problem Size)	Bπ (Background)	Rlen (Rule len)	Rsample (Rule Sample)	Comb. Size
Basic	1	1	5	2	mirror	1	uniform	10³
	2	1	5	2	mirror	1-2	uniform	10³
	3	1	5	4	mirror	1-2	uniform	10⁵
	4	2	5	4	mirror	1-2	uniform	10¹⁰
	5	2	5	6	mirror	1-2	uniform	10¹⁶
Easy	6	2	5	6	uniform	1-2	uniform/llm	10¹⁶
	7	2	6	6	uniform	1-2	uniform/llm	10²⁴
	8	2-3	6	8	uniform	1-2	uniform/llm	10³²
	9	2-3	6	10	uniform	2-3	uniform/llm	10⁴⁰
	10	2-3	7	12	uniform	2-3	uniform/llm	10⁵⁵
Medium	11	2-4	7	14	uniform	2-3	uniform/llm	10⁶⁵
	12	2-4	9	16	uniform	3-4	uniform/llm	10¹²⁰
	13	4-6	9	18	uniform	3-4	uniform/llm	10²⁷¹
	14	4-6	9	20	uniform	4-5	uniform/llm	10³⁰⁰
	15	4-6	9	22	uniform	4-5	uniform/llm	10³³⁰
Hard	16	5-6	10	24	uniform	4-5	uniform/llm	10⁵⁰⁷
	17	5-6	10	26	uniform	4-5	uniform/llm	10⁵⁴⁹
	18	5-6	12	28	uniform	4-5	uniform/llm	10⁸⁰⁵
	19	5-6	12	30	uniform	5	uniform/llm	10⁸⁶¹
	20	5-6	12	32	uniform	5	uniform/llm	10⁹¹⁹

SLR-Bench Curriculum: level-wise configurations, detailing language and task parameters for each difficulty stage. Language complexity is systematically increased by expanding the number of car constants and predicates. Task configuration grows via adapting problem size, background sampling, rule length, and rule sampling strategy. The final column reports the approximate combinatorial size of unique tasks available at each level.

---

Licensing Information

SLR-Bench is made available under the CC BY license.

Citation

If you use this dataset or framework, please cite:

@incollection{helff2025slrautomatedsynthesisscalable,
      title={SLR: Automated Synthesis for Scalable Logical Reasoning}, 
      author={Lukas Helff and Ahmad Omar and Felix Friedrich and Antonia Wüst and Hikaru Shindo and Rupert Mitchell and Tim Woydt and Patrick Schramowski and Wolfgang Stammer and Kristian Kersting},
      year={2025},
      booktitle ={Working Notes of the NeurIPS Workshop on Foundations of Reasoning in Language Models},
      url={https://arxiv.org/abs/2506.15787}, 
}

---