litillabs/octen-law-8b-v1

Octen Law 8B v1

Octen Law 8B v1 is a 7.57B-parameter multilingual dense encoder for legal retrieval. It is built from `Octen/Octen-Embedding-8B` and produces one normalized 4,096-dimensional vector for each query or document.

The release is one merged SentenceTransformers model. It does not require a runtime adapter, reranker, router, retrieval sidecar, or ensemble.

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer(
    "litillabs/octen-law-8b-v1",
    model_kwargs={"torch_dtype": "bfloat16"},
)

queries = model.encode(
    ["Which statute governs this legal issue?"],
    prompt_name="query",
    normalize_embeddings=True,
)
documents = model.encode(
    ["The relevant statutory provision is ..."],
    prompt_name="document",
    normalize_embeddings=True,
)

scores = queries @ documents.T

Training

We post-trained the base model with LoRA on 48,141 public and synthetic English, German, and Chinese legal retrieval examples, then merged the adapter into the model. The checkpoint was selected with held-out retention checks. The eight official MTEB Law scores were not used to choose it.

The training pack included 42,000 rows from public training assets released by Hanno Labs: 20,000 German query-passage pairs and 12,000 Chinese case-pair rows from `Hanno-Labs/legal-retrieval-pairs-v2`, plus 10,000 German Wikipedia query-passage pairs from `Hanno-Labs/legal-retrieval-pairs-dewiki`. Hanno Labs created and released those retrieval pairs; we credit Hanno Labs and Clause Logic Inc. for that work. The remaining 6,141 rows came from LegalPincite, CUAD, and WikiQA training data.

Because Hanno's released training assets make up a substantial part of this model's post-training data and were also used for Dinghy, this result should not be read as an independent training-data comparison with Hanno Labs. A separate independently sourced run is forthcoming.

The training-data audit found six benchmark-adjacent near-duplicate relationships involving GerDaLIRSmall and LeCaRDv2. Those two task scores should not be treated as results from a fully independent holdout.

MTEB Law results

Evaluation used the official test split and cosine similarity for each task.

Limits

This model retrieves text. It does not determine whether an authority is current, binding, overruled, or controlling in the relevant jurisdiction. This release only tested English, German, and Chinese legal retrieval.

License

Apache License 2.0, following the Octen and Qwen3 embedding lineage.