harrier-semantic-v1

Semantic retrieval model for Swiss court decisions — fine-tuned on 55,000 (Sachverhalt → cited decision) pairs.

Understands natural language descriptions of legal situations in German, French, and Italian and retrieves relevant Swiss Federal Court (BGer/BGE) decisions.

Performance (Semantic Benchmark, 800 items)

Metric	Baseline	harrier-semantic-v1
MRR@10	0.0823	0.1501 (+82%)
Recall@10	—	0.0744
Hit@1	—	0.0925

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("ArneH/harrier-semantic-v1")

# Describe a legal situation in natural language
query = "Arbeitnehmer wurde während Krankheit fristlos entlassen"
embedding = model.encode(query, normalize_embeddings=True)

MCP Server (Claude Desktop / Claude Code)

Use with the Swiss Legal Semantic MCP server for offline semantic + keyword search over 963,000+ Swiss court decisions:

pip install mcp sentence-transformers numpy huggingface_hub
python server.py --setup  # auto-downloads model + embeddings

Pre-computed corpus embeddings: ArneH/swiss-caselaw-embeddings

Training

• Base model: harrier-meta-distilled-v2 (0.6B Qwen3-based)
• Training pairs: 55,000 (Sachverhalt → cited decisions, DE/FR/IT)
• Loss: MultipleNegativesRankingLoss
• Epochs: 3, Batch: 512 (2× B200 GPU, DDP)
• Data source: voilaj/swiss-caselaw (963k decisions)