---
title: README
emoji: 🔥
colorFrom: indigo
colorTo: blue
sdk: static
pinned: false
---

# Welcome to HAERAE


We are a non-profit research lab focused on understanding and building better Korean language models. See below for an overview of our projects.

**Benchmarks**  
We have built _the_ most-widely used korean benchmarks including HAE-RAE Bench (cultural knowledge, [dataset](https://huggingface.co/datasets/HAERAE-HUB/HAE_RAE_BENCH_1.0), [paper](https://arxiv.org/abs/2309.02706)), 
KMMLU (general knowledge, [dataset](https://huggingface.co/datasets/HAERAE-HUB/KMMLU), [paper](https://arxiv.org/abs/2402.11548)), HRM8K (math, [dataset](https://huggingface.co/datasets/HAERAE-HUB/HRM8K), [paper](https://www.arxiv.org/abs/2501.02448)), and KMMLU-Redux/Pro  (general knowledge, [dataset](https://huggingface.co/datasets/LGAI-EXAONE/KMMLU-Pro), [paper](https://arxiv.org/abs/2507.08924)).

**Evaluation**  
We developed the [haerae-evaluation-toolkit](https://github.com/HAE-RAE/haerae-evaluation-toolkit), a unified LLM evaluation framework designed to provide consistent and reproducible benchmarking for Korean and multilingual models.

**Reasoning Language Models**  
With cooperation with [KISTI-KONI](https://huggingface.co/KISTI-KONI) we released the [KO-REAson](https://huggingface.co/KOREAson) series, <10B reasoning language models trained for Korean.


# News

2025.08.31: We release six [KO-REAson-0831 models](https://huggingface.co/collections/KoReason/koreason-0831-68b1363e1b3726b041a0a638)  🔥🔥🔥  
2025.07.11: We've collaborated with LG AI Research to build [KMMLU-PRO](https://huggingface.co/datasets/LGAI-EXAONE/KMMLU-Pro) an major update to our KMMLU franchise.  
2025.01.05: We are releasing the first public korean math (📐e = ∑∞ⁿ⁼⁰ ¹ₙ🤓) benchmark [HRM8K](https://huggingface.co/datasets/HAERAE-HUB/HRM8K)