judicialmind/greenleaf-law-embed-tiny

GreenLeaf Law Embed Tiny

by JudicialMind — The AI-Native Legal Intelligence System

A 596M-parameter dense embedding model purpose-built for legal text retrieval. Scores 78.29% on MLEB-12 and 64.38% on MTEB(Law, v1) — competitive with models 13x larger.

Why GreenLeaf?

Most embedding models treat legal text like any other English prose. GreenLeaf doesn't. Every design decision — from bidirectional attention to jurisdiction-aware training data curation — targets the specific challenges of legal retrieval: long documents, precise citations, cross-references, and domain-specific terminology.

Built by JudicialMind as part of our open research initiative to advance legal AI.

Highlights

Parameters	596M
Embedding Size	1024
Context	1024 tokens (supports up to 32,768)
Attention	Bidirectional — every token sees every other token
Pooling	Mean
Precision	bfloat16 (default), int8 and binary quantization built in
License	Apache 2.0

Performance

MLEB-12 — Massive Legal Embedding Benchmark

Dataset	Score
Legal RAG Bench	54.16
Bar Exam QA	68.38
SCALR	73.04
ECHR Retrieval	41.27
Singaporean Judicial Keywords	86.63
GDPR Holdings Retrieval	93.43
UK Legislative Long Titles	95.88
Australian Tax Guidance	78.66
Irish Legislative Summaries	91.92
Contractual Clause Retrieval	91.29
License TL;DR Retrieval	72.54
Consumer Contracts QA	92.29
Overall	78.29

MTEB(Law, v1)

Task	Score
AILACasedocs	40.73
AILAStatutes	58.68
GerDaLIRSmall	38.51
LeCaRDv2	69.52
LegalBenchConsumerContractsQA	85.86
LegalBenchCorporateLobbying	94.94
LegalQuAD	58.62
LegalSummarization	68.22
Overall	64.38

How It Compares

Rank	Model	MLEB	Size
1	Kanon 2 Embedder	81.9%	—
2	Voyage 4 Large	81.1%	7B+
3	Voyage 4	79.6%	—
4	GreenLeaf Law Embed Tiny	78.3%	0.6B
5	Voyage 4 Lite	76.4%	—
6	Qwen3 Embedding 8B	75.9%	8B
7	Gemini Embedding 001	72.1%	—
8	Jina v5 Text Small	71.0%	—
9	OpenAI Text Embedding 3 Large	70.8%	—

Best accuracy-per-parameter on the leaderboard. Runs on a laptop.

Matryoshka Dimension Truncation

GreenLeaf embeddings are trained with Matryoshka Representation Learning — truncate to any dimension without retraining. Halving to 512d costs only 1.34% accuracy; 128d retains 91.6% of full performance.

Dataset	1024d	512d	256d	128d
ECHR Retrieval	40.77	41.35	38.18	33.58
Legal RAG Bench	54.38	49.93	39.98	34.71
SCALR	72.77	70.22	72.50	68.23
Consumer Contracts QA	92.14	92.27	91.13	88.66
Singaporean Judicial Keywords	86.07	84.30	82.63	79.52
Australian Tax Guidance	78.69	77.93	76.22	73.61
Contractual Clause Retrieval	91.20	90.05	87.71	83.31
Irish Legislative Summaries	91.29	90.22	89.51	86.90
GDPR Holdings Retrieval	93.53	93.11	92.32	89.96
License TL;DR Retrieval	73.91	72.83	70.44	71.32
UK Legislative Long Titles	95.88	95.62	96.99	96.73
Bar Exam QA	68.30	65.03	58.79	53.62
MLEB-12 Mean	78.24	76.90	74.70	71.68
Drop vs 1024d	—	-1.34%	-3.54%	-6.56%

# Truncate at inference — no retraining needed
emb_512 = model.encode(texts)[:, :512]
emb_256 = model.encode(texts)[:, :256]

Quick Start

SentenceTransformers

from sentence_transformers import SentenceTransformer

model = SentenceTransformer(
    "judicialmind/greenleaf-law-embed-tiny",
    trust_remote_code=True,
)

texts = [
    "The plaintiff filed a motion for summary judgment.",
    "Contract terms must be interpreted in good faith.",
]

# Default: raw float embeddings (bfloat16)
embeddings = model.encode(texts)  # Shape: (2, 1024)

# Optional: quantize for storage / speed
emb_int8 = model.encode(texts, quantization="int8")
emb_binary = model.encode(texts, quantization="binary")

Text Embeddings Inference (TEI)

docker run --gpus all --shm-size 1g -p 8080:80 \
  ghcr.io/huggingface/text-embeddings-inference:cuda-1.9 \
  --model-id judicialmind/greenleaf-law-embed-tiny --dtype float32

curl http://0.0.0.0:8080/embed \
  -H "Content-Type: application/json" \
  -d '{"inputs": ["Legal text here"], "normalize": false}'

ONNX

import onnxruntime as ort
from transformers import AutoTokenizer
import numpy as np

tokenizer = AutoTokenizer.from_pretrained(
    "judicialmind/greenleaf-law-embed-tiny", trust_remote_code=True
)
session = ort.InferenceSession("onnx/model.onnx")

texts = ["Legal document text here"]
inputs = tokenizer(texts, padding=True, truncation=True, return_tensors="np")
outputs = session.run(
    [o.name for o in session.get_outputs()],
    {"input_ids": inputs["input_ids"].astype(np.int64),
     "attention_mask": inputs["attention_mask"].astype(np.int64)},
)

Architecture

GreenLeaf uses a bidirectional transformer encoder derived from the Qwen3 architecture. The key modification: causal masking is removed across all 28 layers, so every token attends to the full input sequence in both directions.

This matters for legal text. A holding on page 12 can change the meaning of a defined term on page 1. A causal model can't see that. GreenLeaf can.

Property	Value
Base architecture	Qwen3
Hidden layers	28
Hidden size	1024
Attention heads	16 (8 KV heads)
Intermediate size	3072
Vocabulary	151,936
Max position embeddings	32,768
Sliding window	None (full-context)
KV cache	Disabled
Tie embeddings	Yes

Quantization

Three output modes, selectable at inference:

Mode	Output	Storage	Use case
None (default)	float32 per dim	4 KB/doc	Maximum accuracy
"int8"	int8 per dim	1 KB/doc	4x compression, minimal loss
"binary"	±1 per dim	128 B/doc	32x compression, fast Hamming search
"ubinary"	packed bits	128 B/doc	Same as binary, pre-packed

Training

Trained on a large-scale proprietary legal corpus spanning multiple jurisdictions (US, UK, EU, AU, SG, IE) and document types (case law, contracts, legislation, regulatory filings, tax guidance). Contrastive learning with hard negative mining.

About JudicialMind

JudicialMind builds an orchestrated system of specialized AI agents that work like a legal team — every agent with a domain expertise, composed into directed workflows that run end-to-end legal operations: intake, research, drafting, negotiation, dispute resolution.

We serve AmLaw firms, in-house legal teams, solo & boutique practices, and courts/ADR bodies.

Open artifacts:

• 📦 judicialmind/legal-training-dataset — 3.69M annotated query-passage pairs across 35 languages
• 🤖 More models coming soon — retrieval, reranking, and legal-reasoning models trained on our corpus

Connect:

• 🌐 judicialmind.ai
• 📧 surya@judicialmind.ai
• 🧑‍⚖️ Early access

Limitations

• 1024-token default context. Longer documents should be chunked (32k context available with --max-length override).
• Specialist model. For general-purpose embedding, a general-domain model will perform better on non-legal text.

License

Apache 2.0. Free for commercial and research use.