Tool
Pick a Korean dataset by the job, not the name
A long catalog mixes scoring sets with training text. Choose a task below and GearDel keeps rows from this catalog whose license label and original URL are recorded, when those fields exist. Order is not a leaderboard. Files stay with the provider.
What each task means here
Instruction tuning
Rows with an instruction and a response. Not a substitute for a raw crawl.
Benchmarks
Scoring sets such as KLUE, KoBEST, and KMMLU. Do not train on the test items.
Speech recognition
Audio for recognition. Different from read-speech built for synthesis.
Speech synthesis
Closer to text-to-speech than to spontaneous recognition.
OCR
Images with text. Not a generic photo-classification set.
Pretraining text
Longer text such as news, web, or books. It may not be instruction-formatted.
Dialogue
Turns of conversation. Not necessarily an instruction-tuning format.
Sentiment
Sentences with a polarity label. Different from generation instructions.
Commercial use and signup are different filters
“Commercial mark: allowed” only hides rows the catalog does not mark as allowed. “Not AI Hub or NIKL” also hides those two providers, because their downloads usually sit behind an account and a separate agreement. Neither filter is legal advice. If the original LICENSE disagrees with GearDel, follow the original.
After two names survive
Open the compare tool before downloading. It places license, commercial mark, provider, size text, and year side by side. It does not score quality, privacy, or freshness.
Questions
Which Korean dataset should I download for a chatbot?
Split the job first. Instruction rows and dialogue rows can shape replies. Benchmarks such as KLUE, KoBEST, and KMMLU are for scoring, not for training. Use the task buttons on this page, then open the original provider.
Does a commercial-use mark mean I can ship a product?
No. The filter keeps rows the catalog marks as allowed. That is not legal clearance. AI Hub and NIKL items often still need an account. Read the original terms.
Can I download the files here?
No. GearDel does not host dataset files. Each result links to the provider when an original URL is recorded.