opinions config loads the complete slice for 44 jurisdictions (byte-identical to each us-caselaw-<code>): 9,095,360 -> 10,320,499 opinions; counts re-measured 2026-09-17
The single data/*/*.jsonl.gz glob mixed opinions with statutes.v1/rules.v1 files that carry rule_set/rule_number and a per-state-shaped extra struct. HF infers a config's schema from the first file's first batch, so the conversion hit a CastError (column names don't match) and the whole dataset reported 0 rows with viewer:false. Now: one explicit 'opinions' config listing exactly one canonical file set per jurisdiction (no duplicate .v1 snapshots, no statutes/rules), with features declared in dataset_info so every column is a string and extra is a fixed 5-key struct - no inference. Verified with datasets 5.0.1 locally before upload.
credit: Free Law Project / CourtListener as the source of this data
