legal-embed-modernbert-v1

Embedding model for legal question-answer retrieval: given a plain-English legal question, it finds the most relevant answer or passage. Fine-tuned from nomic-ai/modernbert-embed-base on 16K question/answer pairs from Law StackExchange.

Good fit for: legal Q&A search, FAQ retrieval, RAG over legal forums or help-desk content. Supports Matryoshka embeddings — truncate vectors to 512/256/128/64 dims for faster search with modest quality loss.

Evaluation

Scored on 1,000 held-out Law StackExchange Q&A pairs (never seen in training), retrieval among 1,000 candidates:

Model	Accuracy@1	Recall@10	MRR
legal-embed-modernbert-v1	0.893	0.994	0.931
nomic-ai/modernbert-embed-base	0.811	0.955	0.863

Known limitation: on out-of-domain contract-clause retrieval (LegalBench Consumer Contracts QA) this model scores below its base (Accuracy@1 0.556 vs 0.652). It is specialized for conversational legal Q&A, not contract clause search. Use the base model — or wait for v2 — for contract-heavy workloads.

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("bugBug04S/legal-embed-modernbert-v1")

query = "search_query: Can my landlord evict me without notice?"
docs = [
    "search_document: A landlord must generally provide written notice and obtain a court order before eviction...",
    "search_document: Copyright protection begins when a work is fixed in a tangible medium...",
]
q = model.encode(query, normalize_embeddings=True)
d = model.encode(docs, normalize_embeddings=True)
print((q @ d.T))  # higher = more relevant

Important: prepend search_query: to questions and search_document: to passages — the model was trained with these prefixes.

Training

• Base: nomic-ai/modernbert-embed-base (149M params)
• Data: 16,068 (question, top answer) pairs, Law StackExchange (CC-BY-SA 4.0)
• Loss: MatryoshkaLoss(CachedMultipleNegativesRankingLoss), dims 768/512/256/128/64
• 1 epoch, batch 64, lr 1e-5, fp16, max_seq_length 256, single T4 GPU

Disclaimer

Embeddings model for search/retrieval only — it does not generate legal advice. Outputs should not be relied on as a substitute for a qualified lawyer.