FAQ
Korean AI dataset FAQ
Answers for this catalog (286 datasets). Not legal advice.
- Can I download files from GearDel?
- No. GearDel does not store archives. Open the original provider page from the dataset card. If the source URL is unverified, search the provider by the dataset name.
- Is an AI Hub dataset the same as MIT?
- 92 rows in this catalog carry AI Hub terms. Those pages usually need an account and a separate agreement. Putting them in the same table as GitHub MIT is for comparison, not to treat them as Apache 2.0.
- Can I train on KLUE, KoBEST, or KMMLU?
- Those families are scoring sets. Training on the test items raises a leaderboard number. Counts of KLUE, KoBEST, and KMMLU catalog rows are on /en/guides/eval-vs-train. Use the LLM training collection when you need instruction or corpus rows.
- Does “commercial allowed” mean I can ship a product?
- The catalog marks 161 rows as allowed. That is a reading of the recorded license, not legal clearance. Read the original LICENSE. AI Hub and NIKL rows often still need signup.
- Are the size numbers measured on disk?
- No. Size strings are copied from the host catalog text. They can be samples, split archives, or outdated labels.
- What happens if the original page disappears?
- GearDel does not keep a mirror. A dead source should 301 to the home catalog instead of leaving a broken card.
- How do I avoid downloading the wrong Korean dataset?
- Use /en/tools/choose and name the task first. Instruction text, benchmarks, ASR, and TTS are separate filters. /en/compare places license, provider, and size text side by side. Neither page hosts files.
- Can GearDel count tokens or rewrite JSONL?
- Yes, in the browser only: /en/tools/tokens (o200k_base for GPT-4o/GPT-5, cl100k_base for GPT-4; no GPT-6 ranks), /en/tools/jsonl-csv (JSONL to CSV/TSV), /en/tools/cite (catalog-field citations; no invented authors), /en/tools/ko-quality (Hangul mojibake, jamo, mixed speech levels), /en/tools/ko-script (Hangul share, kana, Cyrillic lookalikes), /en/tools/ko-markup (HTML/wiki leftovers), /en/tools/try (catalog excerpts), /en/tools/jsonl (Alpaca, ShareGPT, OpenAI messages), /en/tools/csv-jsonl (spreadsheet to JSONL), /en/tools/jsonl-sample (field viewer), /en/tools/pii (Korean phone, RRN, email patterns), /en/tools/jsonl-split (train/val/test), /en/tools/jsonl-dedup (duplicate lines), /en/tools/jsonl-length (token-length filter), /en/tools/overlap (train vs eval prompts), /en/tools/chat-template (ChatML and Llama 3 wrap), /en/tools/text-stats (UTF-8 and EUC-KR bytes), and /en/tools/hangul (NFC/NFD). The paste is not uploaded.
- How do AI agents get the catalog as JSON?
- Crawl https://geardel.com/llms.txt and https://geardel.com/data/index.json for free. Full records (sourceUrl, samples, search) are x402-paid USDC on Base. Start at GET /api/v1. See /x402.
How to choose → · License comparison → · x402 API → · Token count → · JSONL → · CSV→JSONL → · JSONL split → · JSONL dedup → · Overlap → · PII mask → · JSONL sample → · Data quality → · Try samples → · JSONL to CSV → · Cite a row → · Hangul share → · Scrape leftovers → · Check log → · SFT formats → · Eval vs train →