tojpaj/legal-aid-intake-routing

Legal Aid Intake Routing

Derived from nguha/legalbench (learned_hands_* configs), which are real historical r/legaladvice-style posts labeled by practice area by the LearnedHands legal-aid project. This repackaging keeps only the positively-labeled rows per category and stacks them into a single multi-class classification task.

Task

Given a layperson's plain-language description of a legal problem, predict which practice area it belongs to, so an NGO legal-aid intake system can route it to the right specialist queue instead of a generic backlog ("fast-track resolution").

Files

• train.csv — 4,266 rows (original: text, label)
• test.csv — 755 rows, held out, stratified per label (original: text, label)
• adapted_dataset.csv — the AutoScientist-adapted version of train.csv. Adds enhanced_prompt (a fuller restatement of the situation) and enhanced_completion (a reasoned legal-analysis response), plus row_embedding/row_searchable_text. AutoScientist's own quality evaluation graded this adaptation D → A (score 4.0 → 9.3), an overall +132.5% improvement (message quality +32.9%, completion quality +835.3%) — see autoscientist_evaluation.json.

Columns (train.csv/test.csv): text (intake query), label (one of 16 practice areas: benefits, business, consumer, courts, crime, divorce, domestic_violence, education, employment, estates, family, health, housing, immigration, torts, traffic).

Source & License

Original text and labels: LegalBench / LearnedHands, CC-BY-4.0. See hazyresearch.stanford.edu/legalbench and the LegalBench paper.

Built for the Adaption AutoScientist Challenge, legal category.