brandonrhie/korean-public-figures-saju

Korean Public Figures with Saju (Four Pillars)

39,494 Korean public figures (politicians, actors, athletes, etc.) with verified day-precision birth dates, occupations, and computed Saju (사주, four-pillars / bazi) astrological columns — assembled from Wikidata + Korean Wikipedia, used to empirically test whether Saju can predict occupation.

Dataset Summary

• —Rows: 39,494 (Track A — birth date precision = day)
• —Sources: Wikidata (CC0) seed + Korean Wikipedia (CC BY-SA 3.0) infobox enrichment
• —Saju engine: sajupy 0.2 (Korean Manseryeok / 만세력)
• —Date range: 0995-12-19 ~ 2017-10-28 (most concentrated 1950–1999)
• —Track: A only (year-month-day, no hour pillar — birth-time is missing in <5% of wiki infoboxes, so it was excluded by design)

Use Cases

• —Empirical validation of Saju / Bazi astrology claims (occupation prediction, day-master analysis)
• —Korean public-figure demographics
• —Korean name parsing benchmarks
• —Multi-label occupation classification baseline

Columns

Identity

Birth & Bio

Saju 6 characters (六字, year-month-day stems & branches)

Hour pillar omitted — birth-time is missing for ~95% of public-figure infoboxes, making the fourth pillar unreliable.

Derived (5-elements 오행, yin-yang 음양, 12-zodiac)

Provenance

How the Saju 6-character was computed

For each person, the date triple (year, month, day) is passed to sajupy.calculate_saju(y, m, d, 12, 0). The fixed hour=12, minute=0 choice means the hour pillar from sajupy must be ignored; only the year / month / day pillars are reliable. That is why hour-pillar columns were dropped from the export.

The Manseryeok inside sajupy covers 1900–2100. Persons with birth-year outside that range (mostly pre-1900 historical figures: ~384 rows) get NaN in saju columns.

License

CC BY-SA 4.0 — Creative Commons Attribution-ShareAlike 4.0 International.

This is the strictest license among the source materials used (Wikidata CC0 + Korean Wikipedia CC BY-SA 3.0). If you build on this dataset, you must attribute and share-alike.

Namuwiki data is intentionally excluded from this release, because Namuwiki's CC BY-NC-SA 2.0 KR license would force a non-commercial restriction on the whole dataset.

Attribution

Please cite Wikidata and Korean Wikipedia as the original data sources:

@misc{korean_public_figures_saju_2026,
  title  = {Korean Public Figures with Saju (Four Pillars)},
  year   = {2026},
  note   = {Built from Wikidata (CC0) and Korean Wikipedia dump kowiki-20260501. Saju computation via sajupy.},
  url    = {https://huggingface.co/datasets/brandonrhie/korean-public-figures-saju}
}

Empirical findings (from this dataset)

Using 930 chi-square tests × FDR-corrected (Benjamini-Hochberg α=0.05) across 31 occupations × 15 Saju dimensions, no robust evidence was found that Saju predicts occupation:

• —Day-master (day_stem_element) distribution across the population is nearly uniform (20.3/20.3/19.9/19.9/19.6%), and remains uniform within almost every occupation.
• —The strongest "signals" (Cramér's V ≈ 0.04–0.07) all map onto:
• —Relative Age Effect (RAE) — football/baseball/basketball players cluster on certain month_branch due to school-year cutoff dates, not Saju metaphysics
• —Birth-year cohort effects — politicians cluster on certain year_branch because they are mostly born 1950–1960s
• —All effect sizes are V<0.07 = "negligible" by Cohen's convention.

See the analysis/ files for full reproduction.

Limitations

1. 1.Selection bias: only Korean Wikipedia-notable figures (politicians, athletes, entertainers, etc.). Not generalizable to general population.
2. 2.Hour pillar missing: Saju traditionalists hold that birth-hour is the most important variable. This dataset cannot test that claim because birth-hour is virtually never recorded in public biography sources.
3. 3.Multi-label occupation noise: occupations_raw is preserved as free-form tags (e.g. both "축구 선수" and "대한민국의 축구선수" may appear for the same person). Clustering / normalization required before downstream use.
4. 4.Historical figures: persons born before 1900 are present but their Saju is computed by Manseryeok extrapolation and may have edge cases.

Files

• —train.parquet — main data (Parquet, snappy compression)
• —persons.xlsx — same data in Excel format with column ordering