Dataset Card for XL-AlpacaEval

XL-AlpacaEval is a benchmark for evaluating the cross-lingual open-ended generation capabilities of Large Language Models (LLMs), introduced in the paper XL-Instruct: Synthetic Data for Cross-Lingual Open-Ended Generation. It is designed to evaluate a model's ability to respond in a target language that is different from the source language of the user's query.

For evaluating multilingual (i.e., non-English, but monolingual) generation, see the sister dataset: m-AlpacaEval.

---

🌐 Dataset Details

Focus and Methodology

XL-AlpacaEval is a cross-lingual evaluation benchmark filtered and extended from the AlpacaEval dataset.

• Primary Goal: To assess LLM performance on cross-lingual generation tasks (English-to-X).
• Content: The benchmark contains 797 prompts.
• Evaluation: It uses an LLM-as-a-judge approach, comparing model outputs against a GPT-4o Mini reference, with GPT-4o as the judge.

New languages can be added using the provided add_langs.py script.

• 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), Finnish (fi), Turkish (tr), and Hindi (hi).
• License: CC-BY-SA 4.0.

Sources

• Repository: https://github.com/Remorax/xl-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 cross-lingual generation capabilities of models. It can be used to:

• Benchmark models in the English-to-X direction.
• Track improvements from fine-tuning.
• Analyze performance across different languages.

Out-of-Scope Use

The dataset is not suitable for training models. It is also not designed to evaluate monolingual generation.

---

📊 Dataset Structure

The XL-AlpacaEval dataset consists of 797 multi-domain prompts sourced from the AlpacaEval test set. Each data instance contains:

1. A prompt from the original dataset.
2. An appended cross-lingual instruction (e.g., "Please respond in German.").

Instructions are sampled from a predefined set of templates to add diversity.

---

📈 Evaluation Results

On the XL-AlpacaEval benchmark, most open-weight models in the 7B-9B parameter range demonstrate a significant performance gap compared to the GPT-4o Mini reference model.

• Zero-Shot Performance:

  • Aya Expanse 8B is a top performer, achieving a ~60% win rate in the languages it supports (Portuguese, German, Chinese, Hindi).
  • For most models, performance strongly correlates with the resourcefulness of the target language, with lower scores in low-resource languages.

• Impact of XL-Instruct Fine-Tuning:

  • Fine-tuning on the XL-Instruct dataset significantly boosts performance.
  • Example: Fine-tuning EuroLLM 9B on 8K instructions increases its average win rate from 7.36% to 21.54%.
  • Even small amounts of data are effective for strong models; Aya Expanse 8B improves its win rate from 35.67% to 38.61% with just 32 instructions.

---

🛠️ Dataset Creation

The creation process involved filtering the 805 prompts from AlpacaEval v1 to remove 8 English-centric prompts unsuitable for cross-lingual and multilingual tasks (a full list is provided in the paper's appendix). The remaining 797 prompts were then augmented with templated instructions to generate 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 currently only evaluates the English-to-X direction.

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.