BSC-LT/multi_lmentry

Multi-LMentry

This dataset card provides documentation for Multi-LMentry, a multilingual benchmark designed for evaluating large language models (LLMs) on fundamental, elementary-level tasks across nine languages. It is the official dataset release accompanying the EMNLP 2025 paper "Multi-LMentry: Can Multilingual LLMs Solve Elementary Tasks Across Languages?".

Dataset Details

Dataset Description

Multi-LMentry is a multilingual extension of LMentry (Efrat et al., 2023), which evaluates LLMs on tasks that are trivial for humans but often challenging for models. It covers nine languages:

• Catalan
• German
• Spanish
• Basque
• Galician
• Korean
• Italian
• Brazilian Portuguese
• English (original LMentry dataset)

The dataset enables systematic evaluation of core model abilities across low-, mid-, and high-resource languages. Tasks were recreated manually with the help of native speakers, ensuring linguistic and cultural appropriateness rather than relying on direct translation.

Dataset Sources

• Paper: Accepted at EMNLP 2025 main conference (link pending)
• GitHub Repository: Code to perform the evaluation on Multi-LMentry

Uses

The dataset is intended for:

• Evaluation of LLMs on elementary reasoning and understanding tasks.
• Cross-lingual comparisons, especially between high-resource and low-resource languages.
• Diagnostics / unit tests of fundamental model abilities.

It is not intended for training language models directly.

Dataset Structure

• The dataset is organized by language folders.
• Inside each folder, there is one JSON file per task.
• Each JSON contains input prompts and expected outputs for that task.
• Tasks include simple sentence construction, contextual word choice, alphabetic reasoning, etc.
• Some tasks are language-specific (e.g., rhyming words are excluded where not applicable).

How to Use

from datasets import load_dataset
import json

# Load the Spanish "bigger_number" task
ds = load_dataset(
    "BSC-LT/multi_lmentry",
    "es",
    data_files="es/bigger_number.jsonl"
)["train"]

# Access first example
example = ds[0]
print("Input:", example["input"])

# Convert metadata from string to dictionary
metadata = json.loads(example["metadata"])
print("Metadata:", metadata)

# Access the answer from metadata
answer = metadata.get("answer")
print("Answer:", answer)

Notes:

• The metadata field contains task-specific information, including the answer. Its structure varies depending on the task, for example:
  • Multiple-choice tasks may include a list of distractors and the correct answer index.
  • Open-ended tasks, like "ends_with_letter", may only include task-specific metadata such as the target letter, without a predefined answer.
  • Other fields (e.g., num_digits, n1, n2, template_id) may differ depending on the task type.
• Each JSONL file corresponds to a specific task; you can load multiple tasks by specifying multiple data_files.
• Evaluation: Multi-LMentry includes manually crafted regexes for each task to automatically check answers. These evaluation scripts are available in the (GitHub repository)[https://github.com/langtech-bsc/multi_lmentry] and ready to use for running systematic assessments of model outputs.

Dataset Creation

Curation Rationale

The motivation is to provide a systematic, multilingual benchmark for assessing whether LLMs can perform basic reasoning tasks that humans—even with only elementary proficiency—find trivial. This is crucial since many evaluations today focus on high-level reasoning while overlooking core capabilities.

Source Data

Data Collection and Processing

• Data was manually created in each language, rather than translated from English.
• Native speakers were involved to ensure correctness, cultural relevance, and avoidance of ambiguity or bias.
• Tasks were adapted to respect linguistic characteristics, such as orthography, morphology, or alphabet differences.

Who are the source data producers?

• Native speakers of the target languages, who carefully designed and validated the tasks.
• Task designs follow the original LMentry methodology but were recreated independently per language by native speakers of the target languages, who carefully designed and validated the tasks.

Acknowledgements

Roberto Navigli acknowledges the support, while Simone Conia is fully funded by, the PNRR MUR project PE0000013-FAIR.

The authors gratefully acknowledge the support of the AI Factory IT4LIA project and the CINECA award FAIR_NLP under the ISCRA initiative for granting access high-performance computing resources.

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública and Plan de Recuperación, Transformación y Resiliencia - Funded by EU – NextGenerationEU within the framework of the project ILENIA with references 2022/TL22/00215337, 2022/TL22/00215336 and 2022/TL22/00215335, and within the framework of the project Desarrollo Modelos ALIA.

This work has been promoted and financed by the Generalitat de Catalunya through the Aina project.

License Information

CC BY-SA 4.0

Citation

Bibtex [TODO]