Bge M3 Msmarco V3 Sbert

🚀 BAAI/bge-m3ベースのSentenceTransformer

このモデルは、BAAI/bge-m3 をファインチューニングした sentence-transformers モデルです。文章や段落を1024次元の密ベクトル空間にマッピングし、意味的な文章の類似性、意味検索、パラフレーズマイニング、文章分類、クラスタリングなどに使用できます。

✨ 主な機能

• 文章や段落を1024次元の密ベクトル空間にマッピングすることで、意味的な類似性を計算できます。
• 意味検索、パラフレーズマイニング、文章分類、クラスタリングなどのタスクに利用可能です。

📦 インストール

まず、Sentence Transformersライブラリをインストールします。

pip install -U sentence-transformers

💻 使用例

基本的な使用法

from sentence_transformers import SentenceTransformer

# 🤗 Hubからダウンロード
model = SentenceTransformer("BlackBeenie/bge-m3-msmarco-v3-sbert")
# 推論を実行
sentences = [
    'who is christopher kyle',
    'Chris Kyle American Sniper. Christopher Scott Kyle was born and raised in Texas and was a United States Navy SEAL from 1999 to 2009. He is currently known as the most successful sniper in American military history. According to his book American Sniper, he had 160 confirmed kills (which was from 255 claimed kills).',
    "'American Sniper' Chris Kyle's wife thanks audiences for 'watching the hard stuff'. Taya Kyle has told of her gratitude to audiences for supporting the film about her dead husband Chris Kyle, a Navy Seal played by Bradley Cooper.",
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 1024]

# 埋め込みベクトルの類似度スコアを取得
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]

📚 ドキュメント

モデルの詳細

モデルの説明

モデルのソース

• ドキュメント: Sentence Transformers Documentation
• リポジトリ: Sentence Transformers on GitHub
• Hugging Face: Sentence Transformers on Hugging Face

モデルのアーキテクチャ

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': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
)

学習の詳細

学習データセット

無名データセット

• サイズ: 498,970個の学習サンプル
• 列: <code>sentence_0</code>、<code>sentence_1</code>、<code>sentence_2</code>
• 最初の1000サンプルに基づく概算統計:

  	sentence_0	sentence_1	sentence_2
  タイプ	string	string	string
  詳細	最小: 4トークン 平均: 9.93トークン 最大: 37トークン	最小: 17トークン 平均: 90.01トークン 最大: 239トークン	最小: 16トークン 平均: 86.47トークン 最大: 229トークン

• サンプル:

  sentence_0	sentence_1	sentence_2
  how much does it cost to paint a interior house	Interior House Painting Cost Factors. Generally, it will take a minimum of two gallons of paint to cover a room. At the highest end, paint will cost anywhere between $30 and $60 per gallon and come in three different finishes: flat, semi-gloss or high-gloss.Flat finishes are the least shiny and are best suited for areas requiring frequent cleaning.rovide a few details about your project and receive competitive quotes from local pros. The average national cost to paint a home interior is $1,671, with most homeowners spending between $966 and $2,426.	Question DetailsAsked on 3/12/2014. Guest_... How much does it cost per square foot to paint the interior of a house? We just bought roughly a 1500 sg ft townhouse and want to get the entire house, including ceilings painted (including a roughly 400 sq ft finished basement not included in square footage).
  when is s corp taxes due	If you form a corporate entity for your small business, regardless of whether it's taxed as a C or S corporation, a tax return must be filed with the Internal Revenue Service on its due date each year. Corporate tax returns are always due on the 15th day of the third month following the close of the tax year. The actual day that the tax return filing deadline falls on, however, isn't the same for every corporation.	But if you haven’t, don’t panic: the majority of forms aren’t due quite yet. Most tax forms have an annual January 31 due date. Your tax forms are considered on time if the form is properly addressed and mailed on or before that date. If the regular due date falls on a Saturday, Sunday, or legal holiday – which is the case in 2015 for both January and February due dates – issuers have until the next business day.
  what are disaccharides	Disaccharides are formed when two monosaccharides are joined together and a molecule of water is removed, a process known as dehydration reaction. For example; milk sugar (lactose) is made from glucose and galactose whereas the sugar from sugar cane and sugar beets (sucrose) is made from glucose and fructose.altose, another notable disaccharide, is made up of two glucose molecules. The two monosaccharides are bonded via a dehydration reaction (also called a condensation reaction or dehydration synthesis) that leads to the loss of a molecule of water and formation of a glycosidic bond.	Other disaccharides include (diagrams p. 364): Sucrose, common table sugar, has a glycosidic bond linking the anomeric hydroxyls of glucose and fructose. Because the configuration at the anomeric carbon of glucose is a (O points down from the ring), the linkage is designated a(12).

• 損失関数: MultipleNegativesRankingLoss 、パラメータは以下の通り。

  {
      "scale": 20.0,
      "similarity_fct": "cos_sim"
  }
  

学習ハイパーパラメータ

デフォルトではないハイパーパラメータ

• eval_strategy: steps
• per_device_train_batch_size: 32
• per_device_eval_batch_size: 32
• num_train_epochs: 5
• fp16: True
• multi_dataset_batch_sampler: round_robin

すべてのハイパーパラメータ

学習ログ

フレームワークのバージョン

• Python: 3.11.11
• Sentence Transformers: 3.4.1
• Transformers: 4.48.3
• PyTorch: 2.5.1+cu124
• Accelerate: 1.3.0
• Datasets: 3.3.2
• Tokenizers: 0.21.0

引用

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",
}

MultipleNegativesRankingLoss

@misc{henderson2017efficient,
    title={Efficient Natural Language Response Suggestion for Smart Reply},
    author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil},
    year={2017},
    eprint={1705.00652},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}