hallucite-org/ai-models-eval

Tasks

• text-generation

Citation text from the card

One row per full case citation that eyecite extracted from a brief (short-form citations are not
included), with Hallucite's result.

| column | meaning |
|---|---|
| `brief`, `complaint_id`, `model`, `idx` | which brief, and the citation's position in it |
| `citation` | the citation as written, with its case name |
| `plaintiff`, `defendant` | the case-name parties Hallucite parsed |
| `volume`, `reporter`, `page` | the parsed citation |
| `in_complaint` | true if the complaint itself supplied this citation |
| `in_analysis` | true for the rows counted in the study (see below) |
| `context` | the text surrounding the citation in the brief |
| `status` | Hallucite's result: `exists`, `invalid`, `uncovered`, `incomplete` or `unpublished` |
| `invalid_reason` | why Hallucite judged it invalid, where it did |
| `match_case_name` | the real case Hallucite matched the citation to, if any |
| `plaintiff_match_type`, `defendant_match_type` | how each parsed party compared to the matched case: `exact`, `surname`, `prefix`, `token_subset`, `party_typo`, `surname_typo`, `no_match` |
| `citation_match_type` | `exact`, `volume_typo`, `page_typo`, `typo_2`, `reporter_wrong_series` or `pincite` |
| `is_fuzzy_match`, `name_similarity_score` | whether the name match was fuzzy, and the 0-100 similarity |

**`in_analysis`.** A model is not being tested on a citation the complaint handed it, and
Oklahoma cases are often cited in both neutral and P.3d form, which eyecite emits as two
citations.

All datasets marked Permissive or share-alike | JSON