from src.display.utils import ModelType TITLE = """""" BOTTOM_LOGO = """""" INTRODUCTION_TEXT = f""" 🚀 The Open Ko-LLM Leaderboard 🇰🇷 objectively evaluates the performance of Korean Large Language Model (LLM). When you submit a model on the "Submit here!" page, it is automatically evaluated. The GPU used for evaluation is operated with the support of __[KT](https://cloud.kt.com/)__. The data used for evaluation consists of datasets to assess reasoning, language understanding, hallucination, and commonsense. The evaluation dataset is exclusively private and only available for evaluation process. More detailed information about the benchmark dataset is provided on the “About” page. This leaderboard is co-hosted by __[Upstage](https://www.upstage.ai)__, and __[NIA](https://www.nia.or.kr/site/nia_kor/main.do)__ that provides various Korean Data Sets through __[AI-Hub](https://aihub.or.kr)__, and operated by __[Upstage](https://www.upstage.ai)__. """ LLM_BENCHMARKS_TEXT = f""" # Context While outstanding LLM models are being released competitively, most of them are centered on English and are familiar with the English cultural sphere. We operate the Korean leaderboard, 🚀 Open Ko-LLM, to evaluate models that reflect the characteristics of the Korean language and Korean culture. Through this, we hope that users can conveniently use the leaderboard, participate, and contribute to the advancement of research in Korean. ## Icons {ModelType.PT.to_str(" : ")} model {ModelType.IFT.to_str(" : ")} model {ModelType.RL.to_str(" : ")} model If there is no icon, it indicates that there is insufficient information about the model. Please provide information about the model through an issue! 🤩 🏴‍☠️ : This icon indicates that the model has been selected as a subject of caution by the community, implying that users should exercise restraint when using it. Clicking on the icon will take you to a discussion about that model. (Models that have used the evaluation set for training to achieve a high leaderboard ranking, among others, are selected as subjects of caution.) ## How it works 📈 We evaluate models using the [Eleuther AI Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness), a unified framework to test generative language models on a large number of different evaluation tasks. We have set up a benchmark using datasets translated into Korean, and applied variations by human experts, from the six tasks (HellaSwag, MMLU, Arc, Truthful QA, Winogrande, GSM8k) operated by HuggingFace OpenLLM. We have also added a new dataset prepared from scratch. - Ko-HellaSwag (provided by __[Upstage](https://www.upstage.ai/)__, machine translation) - Ko-MMLU (provided by __[Upstage](https://www.upstage.ai/)__, human translation and variation) - Ko-Arc (provided by __[Upstage](https://www.upstage.ai/)__, human translation and variation) - Ko-Truthful QA (provided by __[Upstage](https://www.upstage.ai/)__, human translation and variation) - Ko-Winogrande (provided by __[Flitto](https://www.flitto.com/portal/en)__, human translation and variation) - Ko-GSM8k (provided by __[Flitto](https://www.flitto.com/portal/en)__, human translation and variation) - Ko-CommonGen V2 (provided by __[Korea University NLP&AI Lab](http://nlp.korea.ac.kr/)__, created from scratch) To provide an evaluation befitting the LLM era, we've selected benchmark datasets suitable for assessing these elements: expertise, inference, hallucination, and common sense. The final score is converted to the average score from each evaluation datasets. GPUs are provided by __[KT](https://cloud.kt.com/)__ for the evaluations. ## Details and Logs - Detailed numerical results in the `results` Upstage dataset: https://huggingface.co/datasets/open-ko-llm-leaderboard/results - Community queries and running status in the `requests` Upstage dataset: https://huggingface.co/datasets/open-ko-llm-leaderboard/requests ## More resources If you still have questions, you can check our FAQ [here](https://huggingface.co/spaces/upstage/open-ko-llm-leaderboard/discussions/1)! """ FAQ_TEXT = """ """ EVALUATION_QUEUE_TEXT = f""" # Evaluation Queue for the 🚀 Open Ko-LLM Leaderboard Models added here will be automatically evaluated on the KT GPU cluster. ## ### 1️⃣ Make sure you can load your model and tokenizer using AutoClasses ```python from transformers import AutoConfig, AutoModel, AutoTokenizer config = AutoConfig.from_pretrained("your model name", revision=revision) model = AutoModel.from_pretrained("your model name", revision=revision) tokenizer = AutoTokenizer.from_pretrained("your model name", revision=revision) ``` If this step fails, follow the error messages to debug your model before submitting it. It's likely your model has been improperly uploaded. ⚠️ Make sure your model is public! ⚠️ Maker sure your model runs with [Eleuther AI Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness) ⚠️ If your model needs `use_remote_code=True`, we do not support this option yet but we are working on adding it, stay posted! ### 2️⃣ Convert your model weights to [safetensors](https://huggingface.co/docs/safetensors/index) It's a new format for storing weights which is safer and faster to load and use. It will also allow us to add the number of parameters of your model to the `Extended Viewer`! ### 3️⃣ Make sure your model has an open license! This is a leaderboard for 🚀 Open Ko-LLMs, and we'd love for as many people as possible to know they can use your model ### 4️⃣ Fill up your model card When we add extra information about models to the leaderboard, it will be automatically taken from the model card ## In case of model failure If your model is displayed in the `FAILED` category, its execution stopped. Make sure you have followed the above steps first. If everything is done, check you can launch the EleutherAIHarness on your model locally, using the above command without modifications (you can add `--limit` to limit the number of examples per task). """ CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results. Authors of open-ko-llm-leaderboard are ordered alphabetically." CITATION_BUTTON_TEXT = r""" @inproceedings{park2024open, title={Open Ko-LLM Leaderboard: Evaluating Large Language Models in Korean with Ko-H5 Benchmark}, author={Chanjun Park and Hyeonwoo Kim and Dahyun Kim and Seonghwan Cho and Sanghoon Kim and Sukyung Lee and Yungi Kim and Hwalsuk Lee}, year={2024}, booktitle={ACL Main} } @software{eval-harness, author = {Gao, Leo and Tow, Jonathan and Biderman, Stella and Black, Sid and DiPofi, Anthony and Foster, Charles and Golding, Laurence and Hsu, Jeffrey and McDonell, Kyle and Muennighoff, Niklas and Phang, Jason and Reynolds, Laria and Tang, Eric and Thite, Anish and Wang, Ben and Wang, Kevin and Zou, Andy}, title = {A framework for few-shot language model evaluation}, month = sep, year = 2021, publisher = {Zenodo}, version = {v0.0.1}, doi = {10.5281/zenodo.5371628}, url = {https://doi.org/10.5281/zenodo.5371628} } @misc{seo2023kocommongen, title={Korean Commonsense Reasoning Evaluation for Large Language Models}, author={Jaehyung Seo, Chanjun Park, Hyeonseok Moon, Sugyeong Eo, Aram So, Heuiseok Lim}, year={2023}, affilation={Korea University, NLP&AI}, booktitle={Proceedings of the 35th Annual Conference on Human & Cognitive Language Technology}} @misc{park2023koarc, title={Ko-ARC}, original_title={Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge}, author={Hyunbyung Park, Chanjun Park}, original_author={Peter Clark and Isaac Cowhey and Oren Etzioni and Tushar Khot and Ashish Sabharwal and Carissa Schoenick and Oyvind Tafjord}, year={2023} } @misc{park2023kohellaswag, title={Ko-HellaSwag}, original_title={HellaSwag: Can a Machine Really Finish Your Sentence?}, author={Hyunbyung Park, Chanjun Park}, original_author={Rowan Zellers and Ari Holtzman and Yonatan Bisk and Ali Farhadi and Yejin Choi}, year={2023} } @misc{park2023kommlu, title={Ko-MMLU}, original_title={Measuring Massive Multitask Language Understanding}, author={Hyunbyung Park, Chanjun Park}, original_author={Dan Hendrycks and Collin Burns and Steven Basart and Andy Zou and Mantas Mazeika and Dawn Song and Jacob Steinhardt}, year={2023} } @misc{park2023kotruthfulqa, title={Ko-TruthfulQA}, original_title={TruthfulQA: Measuring How Models Mimic Human Falsehoods}, author={Hyunbyung Park, Chanjun Park}, original_author={Stephanie Lin and Jacob Hilton and Owain Evans}, year={2023} } """