VERIFICATIONChecked against current official documentation on 2026.08.04; hardware-specific performance is not generalized.
30-SECOND SUMMARY
What to take away
- Create 10–20 Korean prompts with known evidence.
- Fix runtime, prompt, temperature, context, and quantization.
- Score fluency and factual accuracy separately.
KOREAN TEST 01
Compare Korean output
Separate fluency from facts.
- 01SET
Fixed Korean cases
- 02REPEAT
Same conditions
- 03SCORE
Facts, language, format
- 04SELECT
Check critical errors
Break down the work
Summarization, extraction, honorific register, spacing, and structured output are different capabilities. Weight the set according to real use.
Define expected evidence
List required facts and disqualify unsupported claims. Use public or anonymized documents.
Repeat under fixed conditions
Run each case multiple times and record the full model name, version, runtime, and date.
Score separate dimensions
Measure factual match, instruction following, natural Korean, format validity, consistency, and speed.
| Dimension | Check |
|---|---|
| Facts | Names and numbers match |
| Summary | No key omission or invention |
| Language | Consistent terms and register |
| Format | Requested table or JSON is valid |
Frequently asked questions
Does a strong English benchmark guarantee Korean quality?
No. Test the Korean workflow directly.
Is translation testing enough?
No. Include native Korean extraction, summary, and register.
How many repetitions?
Three per prompt is a useful starting point.
Primary sources
Check the original documentation for version-specific details.
Ollama Chat API