You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

To request access to the dataset, please fill out this form, and we'll review and let you know if your use case is approved.

Log in or Sign Up to review the conditions and access this dataset content.

🔴 IMPORTANT❗🔴 HAE-RAE is now LM-Eval supported❗ We are adding evaluation results to our documentation so let us know if you want your model to be added.

The HAE_RAE_BENCH is an ongoing project to develop a suite of evaluation tasks designed to test the understanding of models regarding Korean cultural and contextual nuances. Currently, it comprises 13 distinct tasks, with a total of 4900 instances.

Please note that although this repository contains datasets from the original HAE-RAE BENCH paper, the contents are not completely identical. Specifically, the reading comprehension subset from the original version has been removed due to copyright constraints. In its place, an updated reading comprehension subset has been introduced, sourced from the CSAT, the Korean university entrance examination. To replicate the studies from the paper, please use this code.

For the latest version of the code, refer to this.

Dataset Overview

Task Instances Version Explanation
standard_nomenclature 153 v1.0 Multiple-choice questions about Korean standard nomenclatures from NIKL.
loan_word 169 v1.0 Multiple-choice questions about Korean loan words from NIKL.
rare_word 405 v1.0 Multiple-choice questions about rare Korean words from NIKL.
general_knowledge 176 v1.0 Multiple-choice questions on Korean cultural knowledge.
history 188 v1.0 Multiple-choice questions on Korean history.
date_understanding 475 v1.1 Multiple-choice questions about understanding Korean date expressions.
proverbs_denoising 672 v1.1 Questions for generating Korean proverbs.
correct_definition_matching 439 v1.1 Multiple-choice questions on matching the correct definition to words in context.
lyrics_denoising 622 v1.1 Questions for reconstructing Korean pop song lyrics.
reading_comprehension 936 v1.1 Multiple-choice questions on Korean reading comprehension from the CSAT.
csat_law 217 v1.1 Multiple-choice questions on Korean law from the CSAT.
csat_geo 150 v1.1 Multiple-choice questions on Korean geography from the CSAT.
csat_socio 298 v1.1 Multiple-choice questions on Korean social studies from the CSAT.
Total 4900

Evaluation Results

Models correct_definition_matching csat_geo csat_law csat_socio date_understanding general_knowledge history loan_words reading_comprehension rare_words standard_nomenclature
daekeun-ml/Llama-2-ko-DPO-13B 0.5421 0.1800 0.1613 0.2181 0.4905 0.3523 0.7500 0.8107 0.2382 0.6963 0.7908

How to use (LM-Eval-Harness)

!git clone https://github.com/guijinSON/lm-evaluation-harness.git
!pip install sentencepiece
%cd lm-evaluation-harness
!pip install -e .
!pip install -e ".[multilingual]"
!pip install huggingface_hub
!python -c "from huggingface_hub.hf_api import HfFolder; HfFolder.save_token('<YOUR_HF_TOKEN>')"

!python main.py \
    --model hf-causal \
    --model_args pretrained=daekeun-ml/Llama-2-ko-DPO-13B,dtype=bfloat16\
    --num_fewshot 1\
    --batch_size 2\
    --tasks hr2_cdm,hr2_cgeo,hr2_claw,hr2_csoc,hr2_du,hr2_gk,hr2_hi,hr2_lw,hr2_rw,hr2_rc,hr2_sn\
    --alteration ""\
    --device cuda:0

Release Notes

2023.12.03: All errors fixed! 11 Tasks available via LM-Eval Harness refer to the code above to run evaluation. (List of Available Tasks: correct_definition_matching, csat_geo, csat_law, csat_socio, date_understanding, general_knowledge, history, loan_words, rare_words, reading_comprehension, standard_nomenclature)

2023.11.06: 3 tasks added (csat_geo, csat_law, csat_socio)

2023.09.28: LM-Eval-Harness support added for the following 8 tasks: Loan Words, Rare Words, Standard Nomenclature, History, General Knowledge,correct_definition_matching, date_understanding,reading_comprehension. Refer to the following document to run the evaluation yourself.

2023.09.16: 10 tasks added, 5 from original HAE-RAE Bench(Loan Words, Rare Words, Standard Nomenclature, History, General Knowledge), 5 new tasks (correct_definition_matching, date_understanding, lyrics_denoising, proverbs_denoising, reading_comprehension)

Point of Contact

For any questions contact us via the following email:)

spthsrbwls123@yonsei.ac.kr
Downloads last month
239
Edit dataset card