Skip to content

Benchmark

MLEB - Legal RAG Bench

Embedding benchmarks measure how well models find related items, for example when matching a search to relevant results. Scores are percentages.

Models measured
18
Scores recorded
18
Sources
1
IQ weight
1.06

Score distribution

40.1%
Mean
39.8%
Median
10.5%
Std dev
27%
10th percentile
51.2%
90th percentile
23.8%
Min
23.8%64.8%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.06 for embedding models, measured over 18 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1voyage-4-largeLegal RAG Bench· MLEB data64.8%
2voyage-4Legal RAG Bench· MLEB data56.6%
3voyage-4-liteLegal RAG Bench· MLEB data51.2%
4Qwen3 Embedding 8BLegal RAG Bench· MLEB data50%
5Qwen3 Embedding 4BLegal RAG Bench· MLEB data45.4%
6Jina Embeddings v5 Text SmallLegal RAG Bench· MLEB data42.3%
7Gemini Embedding 001Legal RAG Bench· MLEB data42.2%
8snowflake-arctic-embed-l-v2.0Legal RAG Bench· MLEB data40.2%
9Text Embedding 3 LargeLegal RAG Bench· MLEB data39.8%
10Qwen3 Embedding 0.6BLegal RAG Bench· MLEB data38.7%
11Jina Embeddings v5 Text NanoLegal RAG Bench· MLEB data38.6%
12Jina Embeddings v4Legal RAG Bench· MLEB data37.5%
13Text Embedding 3 SmallLegal RAG Bench· MLEB data37%
14snowflake-arctic-embed-m-v2.0Legal RAG Bench· MLEB data32.8%
15Text Embedding Ada 002Legal RAG Bench· MLEB data27.4%
16Multilingual E5 Large InstructLegal RAG Bench· MLEB data27%
17BGE M3Legal RAG Bench· MLEB data26.8%
18Granite Embedding Small English R2Legal RAG Bench· MLEB data23.8%

Data credits: MLEB (Isaacus), MIT (licence)