HAERAE-en (English-Translated HAERAE-BENCH)

This dataset is the English-translated version of the original HAERAE-BENCH, a benchmark designed to evaluate the linguistic and knowledge-based capabilities of Korean language models. For a detailed understanding of the original dataset's construction and motivation, please refer to the paper: HAE-RAE: A New Public Korean-Specific Benchmark Dataset.

HAERAE-en was created to enable the evaluation of non-Korean models on the knowledge and reasoning aspects of the benchmark, separate from its linguistic challenges. The entire dataset was machine-translated from Korean to English using large language models.

⚠️ Important Disclaimer on Translated Tasks

HAERAE-BENCH contains tasks that are deeply rooted in the nuances of the Korean language. While tasks like History and General Knowledge can be reasonably evaluated in English, the core challenge of the following tasks is significantly altered or diminished through translation:

• Standard Nomenclature (standard_nomenclature): Tests knowledge of standard Korean spelling and usage as defined by the National Institute of Korean Language (NIKL). The translated version may become trivial or lose its original purpose.
• Loan Words (loan_word): Tests knowledge of the correct Korean orthography for words borrowed from other languages. This linguistic test is not applicable in the translated English version.
• Rare Words (rare_word): Tests the understanding of difficult, archaic, or specialized Korean vocabulary. The challenge is fundamentally changed when translated.

Users should be mindful of these limitations when interpreting results from these specific sub-tasks.

Dataset Overview

Task	Instances	Version	Explanation (in the context of the English version)
standard_nomenclature	153	v1.0	(Translated) Multiple-choice questions about standard Korean nomenclatures. Linguistic challenge is lost.
loan_word	169	v1.0	(Translated) Multiple-choice questions about Korean loan words. Linguistic challenge is lost.
rare_word	405	v1.0	(Translated) Multiple-choice questions about rare Korean words. Linguistic challenge is lost.
general_knowledge	176	v1.0	Multiple-choice questions on Korean cultural knowledge.
history	188	v1.0	Multiple-choice questions on Korean history.
reading_comprehension	447	v1.0	Multiple-choice questions on reading comprehension, based on passages from the Korean Language Ability Test (KLAT).
Total	1538

Evaluation Code

To evaluate a model on the HAERAE-en benchmark using EleutherAI's LM-Evaluation-Harness:

# Clone the evaluation harness repository
!git clone [https://github.com/EleutherAI/lm-evaluation-harness.git](https://github.com/EleutherAI/lm-evaluation-harness.git)
!pip install sentencepiece
%cd lm-evaluation-harness
!pip install -e .
!pip install -e ".[multilingual]"
!pip install huggingface_hub

# Run evaluation on the HAERAE-en task
# Note: Using a powerful English-based model as an example
!lm_eval --model hf \
    --model_args pretrained=meta-llama/Llama-3-8B-Instruct \
    --tasks haerae_en \
    --device cuda:0 \
    --batch_size auto:4 \
    --write_out

We've observed minor differences with the original paper's results, which we postulate is mostly due to updates in the lm-evaluation-harness repository.

Point of Contact

For any questions, contact us via the following email:)

[email protected]