README.md

metadata

dataset_info:
  features:
    - name: concept_set
      dtype: string
    - name: '1'
      dtype: string
    - name: '2'
      dtype: string
    - name: '3'
      dtype: string
    - name: '4'
      dtype: string
    - name: gold
      dtype: int64
  splits:
    - name: train
      num_bytes: 1658
      num_examples: 5
    - name: test
      num_bytes: 267519
      num_examples: 847
  download_size: 182862
  dataset_size: 269177
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
task_categories:
  - text-generation
language:
  - ko
pretty_name: KoCommonGEN-V2

🌠 KoCommonGEN v2

KoCommonGEN v2: A Benchmark for Navigating Korean Commonsense Reasoning Challenges in Large Language Models (ACL 2024-Findings)

Jaehyung Seo, Jaewook Lee, Chanjun Park, SeongTae Hong, Seungjun Lee and Heuiseok Lim

🏫 NLP & AI Lab, Korea University

---

🔥 News

• September 27, 2023: Provided data support for the Open Ko-LLM Leaderboard
• August 7, 2024: Dataset Release
• August 10, 2024: Experimental Results for the New Models Added
• August 14, 2024: Presented a research paper at ACL 2024

👥 Human Evaluation

We recruited 22 native Korean speaking volunteers as human evaluators and paid them $0.8 per question.

Model	#	Average Score	cohen's kappa	Krippendorff's alpha
Human	22	0.8395	0.7693	0.7706

🤖 Models (August 10, 2024)

The results of 2-shot evaluation of the newly released models.

Model	Size	Acc_norm	Stderr	Link
GPT-4 (June 13, 2023)		0.7450
Mistral-Nemo-Instruct	12B	0.6612	0.0163	🔗
Mistral-Nemo-Base	12B	0.6340	0.0166	🔗
Meta-Llama-3.1-8B	8B	0.6246	0.0166	🔗
QWEN2-7B base	7B	0.6187	0.0167	🔗
EXAONE-3.0-7.8B-Instruct	7.8B	0.6088	0.0168	🔗
MLP-KTLim-Bllossom-8B	8B	0.6057	0.0168	🔗
Meta-Llama-3.1-8B-Instruct	8B	0.6057	0.0168	🔗
KULLM3	10.8B	0.6033	0.0168	🔗
QWEN2-7B inst	7B	0.5832	0.017	🔗
Gemma-2-9b-it	9B	0.5714	0.0170	🔗
Aya-23-8B	8B	0.5159	0.0172	🔗
Allganize-Alpha-Instruct	8B	0.4970	0.0172	🔗

As mentioned in the paper, it is possible to evaluate various models.

🇰🇷🇺🇸🇯🇵🇨🇳🇪🇸 Code-switching

The dataset can be found on Hugging Face at: nlpai-lab/ko_commongen_v2_code_switching

This dataset contains code-switching data for the following languages:

• Korean (korean)
• English (english)
• Japanese (japan)
• Chinese (china)
• Spanish (espanol)

(The code-switching data relies on machine translation, which may result in some inaccuracies.)

📖 Citation

@inproceedings{seo2024Kocommongenv2,
    title = "KoCommonGEN v2: A Benchmark for Navigating Korean Commonsense Reasoning Challenges in Large Language Models",
    author = "Jaehyung Seo and Jaewook Lee and Chanjun Park and SeongTae Hong and Seungjun Lee and Heuiseok Lim",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2024",
    month = August,
    year = "2024",
    address = "Bangkok, Thailand",
    publisher = "Association for Computational Linguistics",
    url = "TBD",
    doi = "TBD",
    pages = "TBD"}

🚨 Warning!

This dataset contains some instances of toxic speech.

🙏 Acknowledgement

We sincerely appreciate the dedication of Chanjun Park, Sanghoon Kim and Sunghun Kim (Sung Kim) from Upstage AI in managing one of the benchmark datasets for the Open Ko-LLM LeaderBoard.