Dataset Card for m-AlpacaEval

m-AlpacaEval is a benchmark for evaluating the multilingual open-ended generation capabilities of Large Language Models (LLMs) in 8 languages, consisting of Machine Translated prompts from the English XL-AlpacaEval dataset. It was introduced in the paper XL-Instruct: Synthetic Data for Multilingual Open-Ended Generation.

---

🌐 Dataset Details

Focus and Methodology

m-AlpacaEval is a multilingual evaluation benchmark translated from the English XL-AlpacaEval, which in turn is a filtered version from the AlpacaEval test set. The filtered version removes English-specific prompts from AlpacaEval, which we deem cannot be suitably answered in non-English languages (refer to our paper for a full list of excluded prompts).

The prompt translations were produced using the GPT-4o model.

• Primary Goal: To evaluate LLM performance on non-English text generation tasks.
• Content: The benchmark contains 797 prompts machine translated with GPT-4o.
• Evaluation: It employs an LLM-as-a-judge approach, comparing model outputs against a GPT-4o Mini reference, with GPT-4o acting as the judge.

For cross-lingual (English-to-X) evaluation, please see our sister dataset: XL-AlpacaEval.

• Curated by: Vivek Iyer, Ricardo Rei, Pinzhen Chen, and Alexandra Birch (University of Edinburgh, Unbabel)
• Language(s) (NLP): German (de), Portuguese (pt), Hungarian (hu), Lithuanian (lt), Irish (ga), Maltese (mt), Chinese (zh), and Hindi (hi).
• License: CC-BY-NC 4.0.

Sources

• Repository: https://github.com/Remorax/m-alpacaeval
• Paper: XL-Instruct: Synthetic Data for Cross-Lingual Open-Ended Generation

---

🎯 Intended Use

Direct Use

This dataset is intended for direct use as an evaluation benchmark to measure the multilingual generation capabilities of models in non-English languages. It can be used to:

• Benchmark model performance across various languages.
• Track improvements from fine-tuning.
• Analyze monolingual, non-English generation capabilities.

Out-of-Scope Use

The dataset is not suitable for training models. It is also not designed to evaluate cross-lingual generation; for this purpose, please use the XL-AlpacaEval dataset.

---

📊 Dataset Structure

The m-AlpacaEval dataset consists of 797 multi-domain prompts machine translated from the AlpacaEval. Evaluation is to be carried out with LLM-as-a-judge, using https://github.com/tatsu-lab/alpaca_eval evaluation. Code shall be available soon at https://github.com/Remorax/m-alpacaeval.

---

🛠️ Dataset Creation

The creation process involved filtering the 805 prompts from AlpacaEval v1 to remove 8 English-centric prompts unsuitable for multilingual tasks. The remaining 797 prompts were then translated with GPT-4o in the target languages.

• Source Data Producers: The source prompts were originally created by the teams and communities behind OpenAssistant, Koala, Vicuna, Self-Instruct, and Anthropic's Helpfulness test set.

---

⚠️ Bias, Risks, and Limitations

• Inherited Bias: The prompts are derived from English-centric datasets and may contain cultural biases.
• Evaluation Bias: The evaluation relies on GPT-4o as a judge, which may have its own inherent biases.
• Limited Scope: The benchmark evaluates generation in a predefined set of non-English languages.

Recommendations

Users should be aware of these risks and limitations. It is recommended to use this benchmark as one of several tools for a holistic evaluation of a model's capabilities.

---

✍️ Citation

BibTeX

@misc{iyer2025xlinstruct,
title={XL-Instruct: Synthetic Data for Multilingual Open-Ended Generation},
author={Vivek Iyer and Ricardo Rei and Pinzhen Chen and Alexandra Birch},
year={2025},
eprint={2503.22973},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2503.22973},
}

APA

Iyer et al. (2025). XL-Instruct: Synthetic Data for multilingual Open-Ended Generation. arXiv preprint.