Korean DUI Sentencing Dataset: an attorney-verified benchmark of 1,500 cases

A drunk-driving (DUI) subset of the Korean LBox Open judgment corpus, de-duplicated to 1,500 unique cases, recoded to a four-class ordinal sentencing target, and annotated with five sentencing factors that a licensed Korean attorney verified case by case. Each record is one Korean criminal judgment for drunk driving under Article 148-2 of the Road Traffic Act. The target is the imprisonment severity the court imposed, on afour-class ordinal scale: fine or none (215 cases), short, 3-10 months (107), medium, 12-17 months (1,017), and long, 18-30 months (161). The medium class is 67.8% of the data, so any evaluation should report a majority-class baseline and should not rely on accuracy alone. Alongside the judgment text, five legally salient factors are provided as structured columns: blood alcohol concentration, driving distance, vehicle type, number of prior DUIconvictions, and the year of the offence. These were extracted by a frozen rule set and then reviewed in full by a licensed Korean attorney, who corrected 114 of the 7,500extracted values, so the released table is verified rather than automatically labelled. Files are in open formats (JSONL and CSV), with a codebook defining every column. The 1,644-record files reproduce the four source splits exactly and repeat 144 cases; use the1,500-case files for analysis. This is a modified subset of LBox Open (LBox Co., Ltd.), redistributed under CC BY-NC 4.0. The LICENSE file lists the six modifications made. License → Creative Commons Attribution-NonCommercial 4.0 International (cc-by-nc-4.0) columns: blood alcohol concentration, driving distance, vehicle type, number of prior DUI convictions, and the year of the offence. These were extracted by a frozen rule set and then reviewed in full by a licensed Korean attorney, who corrected 114 of the 7,500 extracted values, so the released table is verified rather than automatically labelled. Files are in open formats (JSONL and CSV), with a codebook defining every column. The 1,644-record files reproduce the four source splits exactly and repeat 144 cases; use the 1,500-case files for analysis. This is a modified subset of LBox Open (LBox Co., Ltd.), redistributed under CC BY-NC 4.0. The LICENSE file lists the six modifications made.