KLUE: A Korean NLU Benchmark Dataset

KLUE: A Korean NLU Benchmark Dataset

TL;DR

KLUE is a human-annotated Korean language understanding benchmark built from native Korean sources. It covers eight NLU tasks and includes about 205,000 public train and validation examples.

Now listed on AIOZ AI, KLUE helps developers train and evaluate Korean language models without relying on translated English benchmarks.

What KLUE Is

KLUE is a Korean NLU (Natural Language Understanding) dataset designed for training and evaluating Korean language models across multiple tasks.

Unlike translated benchmark datasets, KLUE is built from native Korean text sources, including news, Wikipedia, product reviews, policy documents, and simulated dialogues. This makes it better suited for modeling how Korean is written, structured, and used in real-world contexts.

For builders working on Korean NLP systems, KLUE provides one consistent benchmark across multiple NLU tasks instead of requiring separate task-specific datasets with different formats and evaluation setups.

How the Dataset Is Organized

KLUE includes eight human-annotated tasks:

  • TC / YNAT: topic classification from news headlines
  • STS: semantic textual similarity
  • NLI: natural language inference
  • NER: named entity recognition
  • RE: relation extraction
  • DP: dependency parsing
  • MRC: machine reading comprehension
  • WoS / DST: dialogue state tracking

Each task has its own schema and annotation format, allowing developers to train or evaluate models on a single capability or benchmark performance across the full Korean NLU stack.

Key Dataset Details

  • Dataset: KLUE
  • Dataset type: Korean NLU benchmark
  • Language: Korean
  • Tasks: 8
  • Public examples: ~205,608 train and validation examples
  • Train split: 171,376 examples
  • Validation split: 34,232 examples
  • Source design: native Korean text, not translated English data
  • Published at: NeurIPS 2021 Datasets and Benchmarks Track
  • License: CC BY-SA 4.0

Where It Fits Best

KLUE is useful for developers, researchers, and teams building Korean language understanding systems.

Practical use cases include:

  • Korean language model evaluation
  • Semantic search and sentence matching
  • Korean named entity recognition for documents and reviews
  • Relation extraction and structured information extraction
  • Full-stack NLU benchmarking before production deployment

Access KLUE on AIOZ AI

KLUE gives builders a consistent benchmark structure to train, evaluate, and compare Korean language models across multiple NLU tasks.

Eight human-annotated tasks, one Korean NLU benchmark.

Access KLUE on AIOZ AI and start building Korean language understanding workflows without starting from scratch.

FAQ

Q1: What is KLUE used for?

KLUE is used to train and evaluate Korean natural language understanding models across eight benchmark tasks.

Q2: How many public examples does KLUE include?

The public train and validation splits include about 205,608 examples.

Q3: What tasks are included in KLUE?

KLUE includes topic classification, semantic similarity, natural language inference, named entity recognition, relation extraction, dependency parsing, machine reading comprehension, and dialogue state tracking.