Litil ColBERT 300M v1

A multilingual legal retrieval model with ColBERT-style late interaction, built by Litil Labs. It is a full-parameter fine-tune of lightonai/mLateOn (revision 35391e36392085d72a93d232f6122607a234ad7a) for German, Chinese and English legal retrieval: case law, statutes, contracts and legal question answering.

307M parameters (mmBERT-base backbone), 128-dimensional token vectors, MaxSim scoring
Query length 1,024 tokens, document length 8,192 tokens, [Q] / [D] prefixes, no query expansion
Runs with PyLate; the checkpoint configuration is authoritative

Evaluation

HAKARI-Bench NanoLaw, exact full-corpus MaxSim, nDCG@10 × 100, FP32.

task	mLateOn (base)	Litil ColBERT 300M v1
NanoLeCaRDv2 (Chinese case retrieval)	62.1	73.1
NanoGerDaLIRSmall (German case retrieval)	54.5	54.9
NanoLegalBenchConsumerContractsQA	88.6	87.8
NanoLegalQuAD (German legal QA)	85.0	84.2
NanoAILACasedocs	26.9	33.0
NanoAILAStatutes	19.8	26.9
NanoLegalBenchCorporateLobbying	94.7	93.0
NanoLegalSummarization	68.0	67.2

Training data

Train splits of public legal retrieval datasets: LeCaRDv2 (Chinese cases), STARD (Chinese statutes), ContractNLI (contract evidence), GerDaLIR (German cases), GerLayQA (German legal QA) and CaseHOLD, with MIRACL and mMARCO as general retention data. Benchmark test queries and documents were excluded from training.

Usage

from pylate import models, rank

model = models.ColBERT(model_name_or_path="litillabs/litil-colbert-300m-v1")

queries = model.encode(["Kündigungsfrist bei befristetem Mietvertrag"], is_query=True)
documents = model.encode(["<document 1>", "<document 2>"], is_query=False)
scores = rank.rerank(
    documents_ids=[[0, 1]],
    queries_embeddings=queries,
    documents_embeddings=[documents],
)

For indexing a corpus, see the PyLate documentation on PLAID indexes. Evaluate in FP32 to reproduce the numbers above.