Legal Case Classifier classifies legal opinions and filings by practice area, procedural posture, outcome, and relevance to a query. Given a case text or summary, it assigns multi-label classifications across 50+ legal categories and procedural postures.

Enables intelligent routing and prioritization of research results in the GLACIER pipeline.

Intended Use

Primary: Classify retrieved case law by practice area and relevance during research
Secondary: Categorize incoming case documents for knowledge base organization
Integration: Powers the research ranking and precedent matching pipeline

Task Type

text-classification -- Multi-label classification across legal taxonomy

Base Model

nlpaueb/legal-bert-base-uncased -- Legal domain pre-training provides strong baseline for legal text understanding

Training Data

Source	Records	Description
CourtListener Opinions	~5M	Court opinions with practice area metadata
Caselaw Access Project	~6.7M	Historical case law with topic classifications
FOLIO Legal Ontology	taxonomy	Legal practice area taxonomy (200+ categories)
Manual Annotations	~100K	Expert-labeled cases for fine-grained categories

Classification Categories (top-level)

Practice areas: tort, contract, property, criminal, family, admin, constitutional, etc.
Procedural posture: MTD, MSJ, trial, appeal, cert, habeas, etc.
Outcome: plaintiff-win, defendant-win, mixed, remanded, reversed, etc.
Jurisdiction: federal, FL state, MS state, other
Relevance: on-point, persuasive, distinguishable, contrary

Benchmark Criteria (90%+ Target)

Metric	Target	Description
Practice Area F1	>= 91%	Multi-label F1 for practice area classification
Procedural Posture Accuracy	>= 93%	Correct procedural posture identification
Outcome Accuracy	>= 90%	Win/loss/mixed classification
Relevance Precision	>= 88%	Must not over-classify as "on-point"
Throughput	>= 100 docs/sec	Batch classification speed

GLACIER Pipeline Integration

STAGE 2 (Research) --> case-classifier filters and categorizes research results
STAGE 3 (WDC #1)  --> classified cases inform jurisdiction/posture validation
STAGE 5 (WDC #2)  --> verify cited cases match the claimed practice area/posture

Training Configuration

Epochs: 8
Learning rate: 2e-5
Batch size: 32
Max sequence length: 512
Loss: Binary cross-entropy (multi-label)
Hardware: AWS SageMaker ml.g5.4xlarge

Limitations

US-focused classification taxonomy; international law categories are sparse
Very recent statutory changes may not be reflected in training data
Edge cases between practice areas may show lower confidence
Does not perform substantive legal analysis, only classification

Version History

Version	Date	Notes
v0.1	2026-04-10	Initial model card, repo created