Natural Innocent
Natural Innocent is a Korean criminal-case dataset for classifying the legal stage at which a charged act was resolved. This public release contains 3,200 case descriptions and their legal-stage labels.
Access is gated. The dataset is released under CC BY 4.0.
Data fields
| Field | Type | Description |
|---|---|---|
id |
string | Stable identifier for the released record. |
source_document_id |
string | Identifier for the source document. |
text |
string | Korean case description. |
label |
integer | Legal-stage label defined below. |
split |
string | One of train, validation, or test. |
decision_year |
string | Year of the source decision. |
Labels
| Label | Name | Description |
|---|---|---|
| 0 | Guilty | The charged conduct resulted in a guilty determination. |
| 1 | Offense elements | One or more offense elements were not established, including the required subjective element. |
| 2 | Unlawfulness | Unlawfulness was negated by a justification or equivalent ground. |
| 3 | Culpability | Criminal responsibility was negated or excused. |
Splits and label distribution
The original study was training-free and used the full dataset for evaluation. The public release additionally provides a deterministic, label-stratified 80/10/10 split for researchers who may need predefined partitions in future work; these partitions were not used as training splits in the original study.
| Split | Label 0 | Label 1 | Label 2 | Label 3 | Total |
|---|---|---|---|---|---|
| Train | 979 | 1,183 | 319 | 79 | 2,560 |
| Validation | 122 | 148 | 40 | 10 | 320 |
| Test | 123 | 148 | 40 | 9 | 320 |
| Total | 1,224 | 1,479 | 399 | 98 | 3,200 |
The 3,200 records correspond to 3,200 distinct source documents. Decision years range from 2004 through 2025.
- Downloads last month
- 12