GLACIER glacier-hallucination-detector

Distilled legal AI model for the GLACIER pipeline — Gated Legal Analysis, Citation Intelligence, Evidence Routing.

Model Description

This model is distilled from Claude Opus 4.6 (via AWS Bedrock) into a lightweight transformer for fast, local inference. It handles citation hallucination detection (real vs fabricated citations) as part of the GLACIER 6-stage legal document production pipeline.

• Base model: distilbert-base-uncased
• Task: sequence-classification
• Labels: 2 classes
• Max length: 128 tokens

Labels

• real
• fabricated

Usage

from glacier_distill.inference import GlacierPipeline

pipeline = GlacierPipeline()
result = pipeline.detect_hallucination("your legal text here")
print(result)

Or use directly with transformers:

from transformers import pipeline

classifier = pipeline("text-classification", model="glacier-legal/glacier-hallucination-detector")
result = classifier("your legal text here")

Training

• Teacher: Claude Opus 4.6 (AWS Bedrock)
• Method: Knowledge distillation (Hinton et al., 2015) with temperature=4.0, alpha=0.7
• Data: CourtListener case law + synthetic labeled examples
• Framework: HuggingFace Transformers + custom DistillationLoss

GLACIER Pipeline

This model is part of the GLACIER pipeline stages:

Stage 1: QUERY    -> jurisdiction-router + document-classifier
Stage 2: RESEARCH -> legal-ner (entity extraction)
Stage 3: WDC #1   -> (full model review)
Stage 4: DRAFT    -> legal-ner + citation-classifier
Stage 5: WDC #2   -> hallucination-detector + citation-classifier
Stage 6: FINAL    -> (human review)

Limitations

• Distilled models are optimized for US legal text (federal + state)
• Not a substitute for full model review in GLACIER Stages 3/5
• Citation hallucination detection is a pre-filter, not a replacement for external verification
• Jurisdiction coverage: Florida, Mississippi, Federal (primary); other states (limited)

License

MIT — Part of the GLACIER Legal AI Framework by Orion Dev Partners, LLC.