Built by merging 5 existing public datasets into one. And I’ve scraped the wiki 69k names<\/a> too.<\/p>\n
| Dataset<\/th>\n | Size<\/th>\n | Male %<\/th>\n | Notes<\/th>\n<\/tr>\n<\/thead>\n |
|---|---|---|---|
| Wikipedia<\/td>\n | 69,209<\/td>\n | 44.1%<\/td>\n | Real attested people, 87% have birth year<\/td>\n<\/tr>\n |
| ENAMDICT<\/td>\n | 116,009<\/td>\n | 16.4%<\/td>\n | Dictionary-based, heavily skewed female<\/td>\n<\/tr>\n |
| Facebook 530M leak<\/td>\n | 392,434<\/td>\n | 60.6%<\/td>\n | Largest source, kanji or kana only<\/td>\n<\/tr>\n |
| GenDec<\/td>\n | 64,139<\/td>\n | 49.8%<\/td>\n | <\/td>\n<\/tr>\n |
| \u540d\u524d\u7531\u6765<\/td>\n | 89,635<\/td>\n | 60.4%<\/td>\n | Popularity rankings, not real frequency<\/td>\n<\/tr>\n |
Total<\/strong><\/td>\n| 731,426<\/strong><\/td>\n | 51.0%<\/strong><\/td>\n | <\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n | Each individual dataset has its own gaps \u2014 size, quality, or skew \u2014 but combining them gives a more complete picture. The Wikipedia subset is the only one covering real individuals and has a temporal dimension through birth years. ENAMDICT skews female partly because Japanese female names have more variety. The Facebook data is massive but only records kanji or<\/em> kana, not both.<\/p>\n Use cases:<\/strong> gender inference (training classifiers without LLMs), Japanese NLP (NER, tokenization, reading prediction), cross-source data quality research<\/p>\n Also working on a gender prediction model, will post when ready. it has around 90% accuracy<\/p>\n<\/div>\n submitted by \/u\/Careful_Sand_7838 <\/a> |
