53,137 five-way multiple-choice questions that ask which holding statement matches a citing passage mined from US judicial opinions.
unassessed
| Category | domain |
|---|---|
| Subcategory | US case-law holding identification |
| Page status | active |
| Metric | macro F1 (paper); exact_match (HELM) |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 53137 |
| Dataset licence | Apache-2.0 |
| Publisher | Stanford RegLab |
CaseHOLD tests whether a model can pick the holding of a cited case from five candidate holding sentences. The prompt is the citing passage from a US opinion. One candidate is the holding that actually follows in that opinion; four are other holdings used as distractors. Holdings are the precedential rule of a decision, so the task is legal citation sense-making, not open legal advice.
Five-way multiple choice. The paper reports mean macro F1 over 10 folds. HELM loads Hugging Face casehold/casehold config `all`, keeps train and test (skips validation), prompts with “Give a letter answer among A, B, C, D, or E,” allows two in-context train items, and scores exact match.
No model card in ModelSpec reports this benchmark yet.