Eval vs train
Korean eval sets vs training rows
GearDel has 286 catalog rows. 47 match the LLM benchmark collection. 63 match the LLM training collection. Those two filters are disjoint: a scoring set is not counted as a training row.
This page does not publish model scores, download totals, or a ranking. License and commercial marks are stored strings. Read the original terms before training or shipping a product.
- KLUE
- 7
- KoBEST
- 5
- KMMLU
- 2
slug family
slug family
slug family
14 rows in those three families. 33 other benchmark rows sit in the same collection (HAE-RAE, CLIcK, and the rest of the catalog filter).
KLUE (7)
Catalog rows whose slug is klue or starts with klue-. How many tasks the original suite has is the size text on each row, not a recount here.
| Name | License | Commercial | Source | Scale text |
|---|---|---|---|---|
| KLUE Benchmark | CC BY-SA 4.0 | Commercial use may be conditional | Hugging Face | 8 · 62.3 MB |
| KLUE | CC BY-SA 4.0 | Commercial use allowed (catalog) | Hugging Face (KLUE) | 12K · Small |
| KLUE | CC BY-SA 4.0 | Commercial use allowed (catalog) | Hugging Face (KLUE) | 25K sentences · Small |
| KLUE | CC BY-SA 4.0 | Commercial use allowed (catalog) | Hugging Face (KLUE) | 11K sentences · Small |
| KLUE | CC BY-SA 4.0 | Commercial use allowed (catalog) | Hugging Face (KLUE) | 21K sentences · Medium |
| KLUE | CC BY-SA 4.0 | Commercial use allowed (catalog) | Hugging Face (KLUE) | 32K sentences · Small |
| KLUE | CC BY-SA 4.0 | Commercial use allowed (catalog) | Hugging Face (KLUE) | 29K items · Medium |
KoBEST (5)
Rows whose slug is kobest or starts with kobest-. An umbrella card and per-task cards can both exist.
| Name | License | Commercial | Source | Scale text |
|---|---|---|---|---|
| KoBEST | CC BY-SA 4.0 | Commercial use allowed (catalog) | Hugging Face | 4.8K items · 5.1MB |
| COPA | CC BY-SA 4.0 | Commercial use allowed (catalog) | Hugging Face (SKT-brain) | 1K items · Small |
| HellaSwag | CC BY-SA 4.0 | Commercial use allowed (catalog) | Hugging Face (SKT-brain) | 2.5K · Small |
| BoolQ | CC BY-SA 4.0 | Commercial use allowed (catalog) | Hugging Face (SKT-brain) | 1.4K items · Small |
| WiC | CC BY-SA 4.0 | Commercial use allowed (catalog) | Hugging Face (SKT-brain) | 3K items · Small |
KMMLU (2)
Only slugs kmmlu and kmmlu-pro. ko-mmlu-pro is a different catalog row.
Ko-MMLU-Pro
A separate slug, not merged into the KMMLU family count above. Open the card for the stored license and source.
What to open instead of a scoring set
Instruction and corpus rows are in Korean LLM training datasets (63 rows). Filter by task on pick. Two names you already know go on compare.
After a local file exists, convert CSV in the browser on CSV→JSONL, check lines on JSONL, sample fields on JSONL sample, mask PII patterns on PII mask, scan Hangul damage on data quality, open catalog excerpts on try samples, flatten JSONL on JSONL to CSV, cite catalog fields on cite, then split on train/val/test. Compare a local train JSONL to pasted eval prompts on train/eval overlap. Format differences are on SFT formats.
Questions
Can I train on KLUE, KoBEST, or KMMLU test items?
Those families are scoring sets in this catalog. Training on the test items raises a leaderboard number. Use the LLM training collection when you need instruction or corpus rows.
Why does the KLUE family row count differ from “8 tasks”?
This page counts 7 catalog rows whose slug is klue or starts with klue-. The “8 tasks” string, if present, is copied from the host on that row. It is not recounted here.
Is Ko-MMLU-Pro the same as KMMLU-Pro?
No. They are separate catalog slugs. Open each card for license and source text.