lebse 0.2.0

LeBSE — legal-domain sentence embeddings adapted from LaBSE

LeBSE — Legal-BERT Sentence Embeddings (a legal-domain LaBSE)

LeBSE adapts Google's LaBSE to the register of U.S. case law, so that legal sentences and opinions land in a better-organized embedding space than the general-purpose base model — while keeping LaBSE's 109-language alignment.

It is a drop-in sentence-transformers model: same API, same 768-dim output, same tokenizer.

from sentence_transformers import SentenceTransformer
m = SentenceTransformer("ahb-sjsu/lebse")          # or a local path
v = m.encode(["The district court lacked subject-matter jurisdiction over the claim."])

Why

General sentence encoders are trained on web text; their geometry is anisotropic (embeddings pile into a narrow cone) and tuned to everyday semantics, not the dense, formulaic register of judicial writing. LeBSE re-tunes the geometry on real opinions so that legally-related text is measurably closer and the space is more isotropic — useful for legal retrieval, clustering, near-duplicate and boilerplate detection, citation recommendation, and as a feature extractor for downstream legal NLP.

How it's trained (v2, citation-supervised)

Citation-supervised contrastive fine-tuning, SPECTER-style: positive pairs are (citing opinion body, cited opinion body) drawn from the CourtListener citation graph, and MultipleNegativesRanking Loss uses the rest of the batch as negatives. Base model: sentence-transformers/LaBSE. Corpus: U.S. federal opinions from the CourtListener / Free Law Project bulk data (public domain). No case outcomes are used.

# lebse-train = SimCSE (v1) baseline; lebse-train-citation = the v2 citation-supervised model.
lebse-train-citation --pairs pairs_train.jsonl --batch 96 --max-seq-len 128 --epochs 2 --out ./lebse

v1 was unsupervised SimCSE and did not beat base LaBSE (it's kept in MODEL_CARD.md for the record). The published weights are v2 (citation-supervised), which does — decisively.

Rigorous evaluation

Two held-out protocols with opinion-level splits (train/eval opinions are disjoint), so any gain is generalization: (1) citation retrieval — the trained relation, on held-out opinions; and (2) docket-lineage retrieval — an independent relation (a district opinion and its appellate reviewer, matched by docket number) the model never trained on. Each reports a paired bootstrap 95% CI on the AUROC gain.

Held-out results (v2, citation-supervised):

eval	base LaBSE	LeBSE-v2	Δ AUROC (95% CI)
citation retrieval (trained relation, held-out)	0.765	0.971	+0.206 [+0.190, +0.223]
docket-lineage (independent relation, unseen)	0.545	0.562	+0.018 [+0.004, +0.031]

LeBSE-v2 dramatically improves citation-type relatedness and transfers a small-but-significant amount to an independent legal relation (district↔appellate lineage) it never trained on. See MODEL_CARD.md for the full protocol and honest caveats. (v1 used unsupervised SimCSE and did not beat LaBSE — kept in the model card for the record.)

LEBSE_DSN="dbname=courtlistener" python eval.py --n 4000 --pairs 3000 --lebse ./lebse

Files

file	purpose
train.py	SimCSE fine-tune of LaBSE on legal sentences
eval.py	held-out, citation-graph-based comparison vs base LaBSE
MODEL_CARD.md	intended use, training data, metrics, limitations
requirements.txt	pinned deps

License

Apache-2.0 (matching the LaBSE base). Training data is U.S. federal case law (public domain). See LICENSE.

Citation

@software{bond_lebse_2026,
  author = {Bond, Andrew H.},
  title  = {LeBSE: Legal-Domain Sentence Embeddings adapted from LaBSE},
  year   = {2026},
  url    = {https://github.com/ahb-sjsu/lebse}
}