SentenceTransformer

This is a sentence-transformers model trained on the train_set dataset. It maps sentences & paragraphs to a 1024-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

Model Details

Learning other languages besides Chinese and English is insufficient, so additional learning is needed to optimize use of other languages.
This model is additionally trained on the Korean dataset.

Model Description

Model Type: Sentence Transformer Transformer Encoder
Maximum Sequence Length: 8192 tokens
Output Dimensionality: 1024 tokens
Similarity Function: Cosine Similarity

Model Sources

Documentation: Sentence Transformers Documentation
Repository: Sentence Transformers on GitHub
Hugging Face: Sentence Transformers on Hugging Face

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 8192, 'do_lower_case': False}) with Transformer model: XLMRobertaModel 
  (1): Pooling({'word_embedding_dimension': 1024, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
  (2): Normalize()
)

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("dragonkue/bge-m3-ko")
# Run inference
sentences = [
    '수급권자 중 근로 능력이 없는 임산부는 몇 종에 해당하니?',
    '내년부터 저소득층 1세 미만 아동의 \n의료비 부담이 더 낮아진다!\n의료급여제도 개요\n□ (목적) 생활유지 능력이 없거나 생활이 어려운 국민들에게 발생하는 질병, 부상, 출산 등에 대해 국가가 의료서비스 제공\n□ (지원대상) 국민기초생활보장 수급권자, 타 법에 의한 수급권자 등\n\n| 구분 | 국민기초생활보장법에 의한 수급권자 | 국민기초생활보장법 이외의 타 법에 의한 수급권자 |\n| --- | --- | --- |\n| 1종 | ○ 국민기초생활보장 수급권자 중 근로능력이 없는 자만으로 구성된 가구 - 18세 미만, 65세 이상 - 4급 이내 장애인 - 임산부, 병역의무이행자 등 | ○ 이재민(재해구호법) ○ 의상자 및 의사자의 유족○ 국내 입양된 18세 미만 아동○ 국가유공자 및 그 유족․가족○ 국가무형문화재 보유자 및 그 가족○ 새터민(북한이탈주민)과 그 가족○ 5․18 민주화운동 관련자 및 그 유가족○ 노숙인 ※ 행려환자 (의료급여법 시행령) |\n| 2종 | ○ 국민기초생활보장 수급권자 중 근로능력이 있는 가구 | - |\n',
    '문재인 대통령의 국정수행 지지율과 더불어민주당의 지지율이 2주 연속 상승했다는 여론조사 결과가 나왔다. 한미정상회담의 효과로 풀이된다. \n한국갤럽은 25~27일 전국 만 18세 이상 1,003명에게 문 대통령의 직무수행 평가를 조사한 결과(표본오차 95% 신뢰수준에 ±3.1%포인트), 37%가 긍정 평가했다고 28일 밝혔다. \n긍정 평가는 지난주보다 3%포인트 올랐다. 부정 평가는 52%로 지난주보다 6%포인트 떨어졌다. 최근 60%대를 넘나들던 부정 평가가 50%대 초반으로 하락했다. 10%는 의견을 유보했다. \n지역별로 보면 서울과 인천·경기의 긍정 평가가 각각 37%란 점이 눈에 띈다. 부산·울산·경남도 33%로, 대구·경북(25%)과 달리 30%대로 나타났다. 연령별로는 40대가 49%로 가장 높았고, 18~29세는 31%로 집계됐다. 정치적 이념·성향이 중도라고 한 응답자의 34%는 긍정 평가했다. \n◇긍정 평가 이유 ‘외교·국제 관계’ 26%P 상승 \n문 대통령의 지지율 상승은 한미정상회담 성과가 영향을 미친 것으로 보인다. 긍정 평가 이유로는 ‘외교·국제 관계’가 가장 높았다. 30%로 지난주보다 26%포인트나 올랐다. 15개월간 ‘신종 코로나바이러스 감염증(코로나19) 대처’가 1위였는데, 한미정상회담 이후 외교·국제 관계로 순위가 바뀌었다. \n다음으로 ‘코로나19 대처’ 22%, ‘최선을 다함·열심히 한다’ 6%, ‘북한 관계’ 4%, ‘전반적으로 잘한다’ 4% 순이었다. \n부정 평가 이유로는 ‘부동산 정책’이 29%로 가장 높았다. 다음으로 ‘경제·민생 문제 해결 부족’ 10%, ‘코로나19 대처 미흡’ 5%, ‘공정하지 못함·내로남불’ 5%, ‘인사 문제’ 4% 순이었다. \n민주당 지지율 역시 문 대통령 지지율과 마찬가지로 2주 연속 상승했다. 정당 지지도 조사에서 민주당은 34%로 지난주보다 2%포인트 올랐다. \n국민의힘은 27%로 지난주보다 1%포인트 올랐다. 민주당과 국민의힘의 지지율 격차는 7%포인트로, 오차범위 밖을 벗어났다. 다음으로 정의당 5%, 열린민주당 3%, 국민의당 3% 순이었다. 무당층은 27%로 조사됐다. \n※자세한 내용은 한국갤럽 또는 중앙선거여론조사심의위원회 홈페이지를 참조하면 된다.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 1024]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]

Evaluation

Metrics

ndcg, mrr, map metrics are metrics that consider ranking, while accuracy, precision, and recall are metrics that do not consider ranking. (Example: When considering ranking for retrieval top 10, different scores are given when the correct document is in 1st place and when it is in 10th place. However, accuracy, precision, and recall scores are the same if they are in the top 10.)

Information Retrieval

Korean Embedding Benchmark is a benchmark with a relatively long 3/4 quantile of string length of 1024

Korean Embedding Benchmark with AutoRAG

This is a benchmark of Korean embedding models. (https://github.com/Marker-Inc-Korea/AutoRAG-example-korean-embedding-benchmark)

Top-k 1

Model name	F1	Recall	Precision	mAP	mRR	NDCG
paraphrase-multilingual-mpnet-base-v2	0.3596	0.3596	0.3596	0.3596	0.3596	0.3596
KoSimCSE-roberta	0.4298	0.4298	0.4298	0.4298	0.4298	0.4298
Cohere embed-multilingual-v3.0	0.3596	0.3596	0.3596	0.3596	0.3596	0.3596
openai ada 002	0.4737	0.4737	0.4737	0.4737	0.4737	0.4737
multilingual-e5-large-instruct	0.4649	0.4649	0.4649	0.4649	0.4649	0.4649
Upstage Embedding	0.6579	0.6579	0.6579	0.6579	0.6579	0.6579
paraphrase-multilingual-MiniLM-L12-v2	0.2982	0.2982	0.2982	0.2982	0.2982	0.2982
openai_embed_3_small	0.5439	0.5439	0.5439	0.5439	0.5439	0.5439
ko-sroberta-multitask	0.4211	0.4211	0.4211	0.4211	0.4211	0.4211
openai_embed_3_large	0.6053	0.6053	0.6053	0.6053	0.6053	0.6053
KU-HIAI-ONTHEIT-large-v1	0.7105	0.7105	0.7105	0.7105	0.7105	0.7105
KU-HIAI-ONTHEIT-large-v1.1	0.7193	0.7193	0.7193	0.7193	0.7193	0.7193
kf-deberta-multitask	0.4561	0.4561	0.4561	0.4561	0.4561	0.4561
gte-multilingual-base	0.5877	0.5877	0.5877	0.5877	0.5877	0.5877
BGE-m3	0.6578	0.6578	0.6578	0.6578	0.6578	0.6578
BGE-m3-ko	0.7456	0.7456	0.7456	0.7456	0.7456	0.7456

Top-k 3

Model name	F1	Recall	Precision	mAP	mRR	NDCG
paraphrase-multilingual-mpnet-base-v2	0.2368	0.4737	0.1579	0.2032	0.2032	0.2712
KoSimCSE-roberta	0.3026	0.6053	0.2018	0.2661	0.2661	0.3515
Cohere embed-multilingual-v3.0	0.2851	0.5702	0.1901	0.2515	0.2515	0.3321
openai ada 002	0.3553	0.7105	0.2368	0.3202	0.3202	0.4186
multilingual-e5-large-instruct	0.3333	0.6667	0.2222	0.2909	0.2909	0.3856
Upstage Embedding	0.4211	0.8421	0.2807	0.3509	0.3509	0.4743
paraphrase-multilingual-MiniLM-L12-v2	0.2061	0.4123	0.1374	0.1740	0.1740	0.2340
openai_embed_3_small	0.3640	0.7281	0.2427	0.3026	0.3026	0.4097
ko-sroberta-multitask	0.2939	0.5877	0.1959	0.2500	0.2500	0.3351
openai_embed_3_large	0.3947	0.7895	0.2632	0.3348	0.3348	0.4491
KU-HIAI-ONTHEIT-large-v1	0.4386	0.8772	0.2924	0.3421	0.3421	0.4766
KU-HIAI-ONTHEIT-large-v1.1	0.4430	0.8860	0.2953	0.3406	0.3406	0.4778
kf-deberta-multitask	0.3158	0.6316	0.2105	0.2792	0.2792	0.3679
gte-multilingual-base	0.4035	0.8070	0.2690	0.3450	0.3450	0.4614
BGE-m3	0.4254	0.8508	0.2836	0.3421	0.3421	0.4701
BGE-m3-ko	0.4517	0.9035	0.3011	0.3494	0.3494	0.4886

Top-k 5

Model name	F1	Recall	Precision	mAP	mRR	NDCG
paraphrase-multilingual-mpnet-base-v2	0.1813	0.5439	0.1088	0.1575	0.1575	0.2491
KoSimCSE-roberta	0.2164	0.6491	0.1298	0.1751	0.1751	0.2873
Cohere embed-multilingual-v3.0	0.2076	0.6228	0.1246	0.1640	0.1640	0.2731
openai ada 002	0.2602	0.7807	0.1561	0.2139	0.2139	0.3486
multilingual-e5-large-instruct	0.2544	0.7632	0.1526	0.2194	0.2194	0.3487
Upstage Embedding	0.2982	0.8947	0.1789	0.2237	0.2237	0.3822
paraphrase-multilingual-MiniLM-L12-v2	0.1637	0.4912	0.0982	0.1437	0.1437	0.2264
openai_embed_3_small	0.2690	0.8070	0.1614	0.2148	0.2148	0.3553
ko-sroberta-multitask	0.2164	0.6491	0.1298	0.1697	0.1697	0.2835
openai_embed_3_large	0.2807	0.8421	0.1684	0.2088	0.2088	0.3586
KU-HIAI-ONTHEIT-large-v1	0.3041	0.9123	0.1825	0.2137	0.2137	0.3783
KU-HIAI-ONTHEIT-large-v1.1	0.3099	0.9298	0.1860	0.2148	0.2148	0.3834
kf-deberta-multitask	0.2281	0.6842	0.1368	0.1724	0.1724	0.2939
gte-multilingual-base	0.2865	0.8596	0.1719	0.2096	0.2096	0.3637
BGE-m3	0.4254	0.8508	0.2836	0.3421	0.3421	0.4701
BGE-m3-ko	0.3099	0.9298	0.1860	0.2098	0.2098	0.3793

Top-k 10

Model name	F1	Recall	Precision	mAP	mRR	NDCG
paraphrase-multilingual-mpnet-base-v2	0.1212	0.6667	0.0667	0.1197	0.1197	0.2382
KoSimCSE-roberta	0.1324	0.7281	0.0728	0.1080	0.1080	0.2411
Cohere embed-multilingual-v3.0	0.1324	0.7281	0.0728	0.1150	0.1150	0.2473
openai ada 002	0.1563	0.8596	0.0860	0.1051	0.1051	0.2673
multilingual-e5-large-instruct	0.1483	0.8158	0.0816	0.0980	0.0980	0.2520
Upstage Embedding	0.1707	0.9386	0.0939	0.1078	0.1078	0.2848
paraphrase-multilingual-MiniLM-L12-v2	0.1053	0.5789	0.0579	0.0961	0.0961	0.2006
openai_embed_3_small	0.1547	0.8509	0.0851	0.0984	0.0984	0.2593
ko-sroberta-multitask	0.1276	0.7018	0.0702	0.0986	0.0986	0.2275
openai_embed_3_large	0.1643	0.9035	0.0904	0.1180	0.1180	0.2855
KU-HIAI-ONTHEIT-large-v1	0.1707	0.9386	0.0939	0.1105	0.1105	0.2860
KU-HIAI-ONTHEIT-large-v1.1	0.1722	0.9474	0.0947	0.1033	0.1033	0.2822
kf-deberta-multitask	0.1388	0.7632	0.0763	0.1	0.1	0.2422
gte-multilingual-base	0.1675	0.9211	0.0921	0.1066	0.1066	0.2805
BGE-m3	0.4254	0.8508	0.2836	0.3421	0.3421	0.4701
BGE-m3-ko	0.1770	0.9736	0.0974	0.1097	0.1097	0.2932

Information Retrieval

Dataset: miracl-ko (https://github.com/project-miracl/miracl)
miracl benchmark is a benchmark with a relatively short 3/4 quantile of string length of 220 on the Korean Wikidata set.
Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6103
cosine_accuracy@3	0.8169
cosine_accuracy@5	0.8732
cosine_accuracy@10	0.9202
cosine_precision@1	0.6103
cosine_precision@3	0.3787
cosine_precision@5	0.2761
cosine_precision@10	0.1728
cosine_recall@1	0.3847
cosine_recall@3	0.5902
cosine_recall@5	0.6794
cosine_recall@10	0.7695
cosine_ndcg@10	0.6723
cosine_mrr@10	0.7262
cosine_map@100	0.6074
dot_accuracy@1	0.6103
dot_accuracy@3	0.8169
dot_accuracy@5	0.8732
dot_accuracy@10	0.9202
dot_precision@1	0.6103
dot_precision@3	0.3787
dot_precision@5	0.2761
dot_precision@10	0.1728
dot_recall@1	0.3847
dot_recall@3	0.5902
dot_recall@5	0.6794
dot_recall@10	0.7695
dot_ndcg@10	0.6723
dot_mrr@10	0.7262
dot_map@100	0.6074

Bias, Risks and Limitations

Since the evaluation results are different for each domain, it is necessary to compare and evaluate the model in your own domain. In the Miracl benchmark, the evaluation was conducted using the Korean Wikipedia as a corpus, and in this case, the cosine_ndcg@10 score dropped by 0.2 points after learning. However, in the Auto-RAG benchmark, which is a financial domain, the ndcg score increased by 0.9 when it was top 1. This model may be advantageous for use in a specific domain.
Also, since the miracl benchmark consists of a corpus of relatively short strings, while the Korean Embedding Benchmark consists of a corpus of longer strings, this model may be more advantageous if the length of the corpus you want to use is long.

Training Hyperparameters

Non-Default Hyperparameters

The batch size was referenced from the following paper: Text Embeddings by Weakly-Supervised Contrastive Pre-training (https://arxiv.org/pdf/2212.03533)

eval_strategy: steps
per_device_train_batch_size: 32768
per_device_eval_batch_size: 32768
learning_rate: 3e-05
warmup_ratio: 0.03333333333333333
fp16: True
batch_sampler: no_duplicates

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: steps
prediction_loss_only: True
per_device_train_batch_size: 32768
per_device_eval_batch_size: 32768
per_gpu_train_batch_size: None
per_gpu_eval_batch_size: None
gradient_accumulation_steps: 1
eval_accumulation_steps: None
learning_rate: 3e-05
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
max_grad_norm: 1.0
num_train_epochs: 3
max_steps: -1
lr_scheduler_type: linear
lr_scheduler_kwargs: {}
warmup_ratio: 0.03333333333333333
warmup_steps: 0
log_level: passive
log_level_replica: warning
log_on_each_node: True
logging_nan_inf_filter: True
save_safetensors: True
save_on_each_node: False
save_only_model: False
restore_callback_states_from_checkpoint: False
no_cuda: False
use_cpu: False
use_mps_device: False
seed: 42
data_seed: None
jit_mode_eval: False
use_ipex: False
bf16: False
fp16: True
fp16_opt_level: O1
half_precision_backend: auto
bf16_full_eval: False
fp16_full_eval: False
tf32: None
local_rank: 0
ddp_backend: None
tpu_num_cores: None
tpu_metrics_debug: False
debug: []
dataloader_drop_last: True
dataloader_num_workers: 0
dataloader_prefetch_factor: None
past_index: -1
disable_tqdm: False
remove_unused_columns: True
label_names: None
load_best_model_at_end: False
ignore_data_skip: False
fsdp: []
fsdp_min_num_params: 0
fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
fsdp_transformer_layer_cls_to_wrap: None
accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
deepspeed: None
label_smoothing_factor: 0.0
optim: adamw_torch
optim_args: None
adafactor: False
group_by_length: False
length_column_name: length
ddp_find_unused_parameters: None
ddp_bucket_cap_mb: None
ddp_broadcast_buffers: False
dataloader_pin_memory: True
dataloader_persistent_workers: False
skip_memory_metrics: True
use_legacy_prediction_loop: False
push_to_hub: False
resume_from_checkpoint: None
hub_model_id: None
hub_strategy: every_save
hub_private_repo: False
hub_always_push: False
gradient_checkpointing: False
gradient_checkpointing_kwargs: None
include_inputs_for_metrics: False
eval_do_concat_batches: True
fp16_backend: auto
push_to_hub_model_id: None
push_to_hub_organization: None
mp_parameters:
auto_find_batch_size: False
full_determinism: False
torchdynamo: None
ray_scope: last
ddp_timeout: 1800
torch_compile: False
torch_compile_backend: None
torch_compile_mode: None
dispatch_batches: None
split_batches: None
include_tokens_per_second: False
include_num_input_tokens_seen: False
neftune_noise_alpha: None
optim_target_modules: None
batch_eval_metrics: False
batch_sampler: no_duplicates
multi_dataset_batch_sampler: proportional

Citation

BibTeX

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

@misc{bge-m3,
      title={BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation}, 
      author={Jianlv Chen and Shitao Xiao and Peitian Zhang and Kun Luo and Defu Lian and Zheng Liu},
      year={2024},
      eprint={2402.03216},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@article{wang2022text,
  title={Text Embeddings by Weakly-Supervised Contrastive Pre-training},
  author={Wang, Liang and Yang, Nan and Huang, Xiaolong and Jiao, Binxing and Yang, Linjun and Jiang, Daxin and Majumder, Rangan and Wei, Furu},
  journal={arXiv preprint arXiv:2212.03533},
  year={2022}
}

dragonkue
/

bge-m3-ko

SentenceTransformer

Model Details

Model Description

Model Sources

Full Model Architecture

Usage

Direct Usage (Sentence Transformers)

Evaluation

Metrics

Information Retrieval

Korean Embedding Benchmark with AutoRAG

Information Retrieval

Bias, Risks and Limitations

Training Hyperparameters

Non-Default Hyperparameters

All Hyperparameters

Citation

BibTeX

Sentence Transformers

Model tree for dragonkue/bge-m3-ko

Evaluation results