Benchmark for Japanese document embedding and vector search
requirements:
- python >=3.9,<3.12
- poetry
required environmental variables:
OPENAI_API_KEY
(to use OpenAI's embedding API)COHERE_API_KEY
(to use Cohere's embedding API)
optional:
DBSV_CACHE_DIR
(for cache directory, default:~/.dvsb
)
to install all dependencies, perform
$ make install
Benchmark settings can be controlled through a yaml file. The default configuration can be found in configs/default.yml.
You can use your own config file in the configs directry by setting the DVSB_CONFIG_NAME
environmental variable (ex. default
).
To run benchmark,
$ make run_benchmark
Model | Recall@1 JSQuAD-v1.1-valid | Recall@3 JSQuAD-v1.1-valid | Recall@5 JSQuAD-v1.1-valid | Recall@10 JSQuAD-v1.1-valid | Recall@3 MIRACL-v1.0-dev | Recall@5 MIRACL-v1.0-dev | Recall@10 MIRACL-v1.0-dev | Recall@100 MIRACL-v1.0-dev | |
---|---|---|---|---|---|---|---|---|---|
0 | ColBERTRetriever-bclavie/JaColBERTv2 | 0.920756 | 0.967807 | 0.976587 | 0.982665 | 0.622286 | 0.713644 | 0.82424 | 0.970855 |
1 | SentenceTransformerEmbedding-BAAI/bge-m3 | 0.849842 | 0.939217 | 0.958577 | 0.975912 | 0.686322 | 0.769516 | 0.85376 | 0.980984 |
2 | E5Embedding-intfloat/multilingual-e5-large | 0.864926 | 0.952949 | 0.965781 | 0.977488 | 0.658555 | 0.741041 | 0.835273 | 0.982559 |
3 | ColBERTRetriever-bclavie/JaColBERT | 0.911301 | 0.961054 | 0.970059 | 0.977262 | 0.555464 | 0.639378 | 0.748628 | 0.933103 |
4 | E5Embedding-intfloat/multilingual-e5-base | 0.838361 | 0.934039 | 0.954975 | 0.972535 | 0.612025 | 0.687829 | 0.798884 | 0.976868 |
5 | E5Embedding-intfloat/multilingual-e5-small | 0.840387 | 0.933814 | 0.95385 | 0.972985 | 0.59911 | 0.688636 | 0.783391 | 0.972148 |
6 | VertexAITextEmbedding-textembedding-gecko-multilingual@001 | 0.780729 | 0.904322 | 0.932463 | 0.961054 | N/A | N/A | N/A | N/A |
7 | VertexAITextEmbedding-textembedding-gecko-multilingual@latest | 0.780729 | 0.904548 | 0.932238 | 0.960603 | N/A | N/A | N/A | N/A |
8 | OpenAIEmbedding-text-embedding-ada-002 | 0.75394 | 0.874606 | 0.906799 | 0.937866 | N/A | N/A | N/A | N/A |
9 | SonoisaSentenceLukeJapanese-sonoisa/sentence-luke-japanese-base-lite | 0.652634 | 0.813147 | 0.861324 | 0.908825 | 0.144617 | 0.211484 | 0.297427 | 0.622732 |
10 | SonoisaSentenceBertJapanese-sonoisa/sentence-bert-base-ja-mean-tokens-v2 | 0.65421 | 0.810671 | 0.8629 | 0.914228 | 0.170388 | 0.221441 | 0.314252 | 0.660344 |
11 | SentenceTransformerEmbedding-pkshatech/GLuCoSE-base-ja | 0.644755 | 0.798064 | 0.846466 | 0.896668 | 0.472039 | 0.546962 | 0.645746 | 0.861757 |
12 | SentenceTransformerEmbedding-cl-nagoya/sup-simcse-ja-base | 0.631923 | 0.792661 | 0.848942 | 0.897118 | 0.136905 | 0.185814 | 0.26734 | 0.590408 |
13 | SentenceTransformerEmbedding-sonoisa/sentence-bert-base-ja-mean-tokens-v2 | 0.639802 | 0.782981 | 0.841288 | 0.894867 | 0.203452 | 0.26437 | 0.357611 | 0.708596 |
14 | SentenceTransformerEmbedding-cl-nagoya/sup-simcse-ja-large | 0.603107 | 0.776452 | 0.833408 | 0.889239 | 0.154618 | 0.202999 | 0.293636 | 0.584585 |
15 | SentenceTransformerEmbedding-cl-nagoya/unsup-simcse-ja-large | 0.594777 | 0.755966 | 0.8181 | 0.879559 | 0.102252 | 0.142104 | 0.218686 | 0.52806 |
16 | SentenceTransformerEmbedding-cl-nagoya/unsup-simcse-ja-base | 0.577217 | 0.746961 | 0.804142 | 0.870779 | 0.0963266 | 0.121559 | 0.195299 | 0.500001 |