SauerkrautLM-MoEdernColBERT-Legal
Collection
German legal ColBERT retrievers on the MoEdernBERT architecture (1.5M, 27M, 115M) with benchmark, leaderboard and demo • 6 items • Updated • 2
SauerkrautLM-MoEdernColBERT-1.5M-Legal-DE
German legal retrieval with 1.5M parameters: 0.388 on OpenLegalRetrieval-DE, ahead of bge-multilingual-gemma2 (0.377) at 1/6200 the size.
German legal retrieval for statutes, case law and legal questions, specialized from SauerkrautLM-MoEdernColBERT-1.5M. MoEdernBERT architecture: 1.47M parameters, 0.88M active per token, ±256-token sliding window for long documents.
Family: 1.5M · 27M · 115M · Collection
Highlights
- LegalQuAD 0.813: ahead of 8B models and commercial APIs (dinghy-law-8b-v1 0.773, voyage-4-large 0.769†).
- GerDaLIRSmall 0.462: ahead of Qwen3-Embedding-8B (0.454) at 1/5100 the size.
† public MTEB result, not measured by us.
Results
Usage
pip install -U "sentence-transformers>=6.1"
pip install -U triton # optional: Linux + NVIDIA GPU
from sentence_transformers import MultiVectorEncoder
model = MultiVectorEncoder(
"VAGOsolutions/SauerkrautLM-MoEdernColBERT-1.5M-Legal-DE",
trust_remote_code=True,
)
queries = model.encode_query([
"Wie lange ist die regelmäßige Verjährungsfrist?",
])
documents = model.encode_document([
"§ 195 BGB: Die regelmäßige Verjährungsfrist beträgt drei Jahre.",
"§ 433 BGB: Durch den Kaufvertrag wird der Verkäufer ...",
])
print(model.similarity(queries, documents)) # MaxSim
- Shipped configuration: documents up to 8,192 tokens, queries up to 2,048 tokens (query mode
min), ±256-token window. - Fast GPU path: on CUDA with Triton installed, the bundled fused Triton MoE kernel is used automatically for inference;
MOEDERN_TRITON=0turns it off. - Precision: fp32 or bf16, not fp16. PyLate has no query mode
min: setquery_length=32.
Evaluation
Own measurements on the same data for every model, each with its own configuration; † public MTEB result (read 2026-11-03); ⚠ GerDaLIR/GerDaLIRSmall in the model's training data per MTEB metadata. GerDaLIRSmall is full7 task T7 with standard query encoding. No test queries or qrels of these benchmarks were used in training.
German Legal Embedding Leaderboard · code and submissions · OpenLegalRetrieval-DE · demo: legal search
License
CC BY-NC 4.0 for research and other non-commercial use.
Citation
@misc{sauerkrautlm_moederncolbert_1_5m_legal_de_2026,
title={SauerkrautLM-MoEdernColBERT-1.5M-Legal-DE},
author={David Golchinfar},
organization={VAGO Solutions},
url={https://huggingface.co/VAGOsolutions/SauerkrautLM-MoEdernColBERT-1.5M-Legal-DE},
year={2026}
}
- Downloads last month
- 37
from sentence_transformers import MultiVectorEncoder model = MultiVectorEncoder("VAGOsolutions/SauerkrautLM-MoEdernColBERT-1.5M-Legal-DE", trust_remote_code=True) queries = ["Which planet is known as the Red Planet?"] documents = [ "Venus is often called Earth's twin because of its similar size and proximity.", "Mars, known for its reddish appearance, is often referred to as the Red Planet.", "Jupiter, the largest planet in our solar system, has a prominent red spot.", ] query_embeddings = model.encode_query(queries) document_embeddings = model.encode_document(documents) similarities = model.similarity(query_embeddings, document_embeddings) print(similarities)