Acgvoc2vec

🚀 ACGVoc2vec

ACGVoc2vec is a model structured on sentence-transformers. It fine-tunes the pre - trained weights of distiluse-base-multilingual-cased-v2 on an anime-related sentence-pair dataset with a learning rate of 5e-5, using MultipleNegativesRankingLoss as the loss function.

Pipeline and Tags

Widget

• Source Sentence: "亚丝娜"
• Comparison Sentences: ["火影忍者", "Sword Art Online", "结城明日奈", "アスナ"]

🚀 Quick Start

📦 Installation

Using this model becomes easy when you have sentence-transformers installed:

pip install -U sentence-transformers

💻 Usage Examples

Basic Usage

from sentence_transformers import SentenceTransformer
sentences = ["This is an example sentence", "Each sentence is converted"]

model = SentenceTransformer('OysterQAQ/ACGVoc2vec')
embeddings = model.encode(sentences)
print(embeddings)

🔧 Technical Details

The model is structured as sentence-transformers. It uses the pre - trained weights of distiluse-base-multilingual-cased-v2 and fine - tunes them on an anime - related sentence - pair dataset with a learning rate of 5e - 5. The loss function is MultipleNegativesRankingLoss.

Dataset

The dataset mainly includes:

• Bangumi:
  • Japanese anime names - Chinese anime names
  • Japanese anime names - Summaries
  • Chinese anime names - Summaries
  • Chinese anime names - Tags
  • Japanese anime names - Characters
  • Chinese anime names - Characters
  • Japanese voice actor names - Chinese voice actor names
• pixiv:
  • Japanese tag names - Chinese tag names
• AnimeList:
  • Japanese anime names - English anime names
• Wikipedia:
  • Japanese anime names - Chinese anime names
  • Japanese anime names - English anime names
  • h2 titles and corresponding texts on Chinese, English, and Japanese detail pages
  • Multilingual summaries (Chinese, Japanese, English)
  • Anime names - Summaries (Chinese, Japanese, English)
• moegirl:
  • Summaries of Chinese anime names - Summaries
  • Chinese anime names + Sub - titles - Corresponding content

After crawling, cleaning, and processing, 80 million text pairs are obtained (and the number is still increasing). The model is trained for 20 epochs with a batch size of 80, enabling the weights of sentence - transformers to adapt to this problem space and generate text feature vectors that incorporate domain knowledge (manifested as closer distances between relevant texts, such as works and characters, or characters from the same work).

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 128, 'do_lower_case': False}) with Transformer model: DistilBertModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False})
  (2): Dense({'in_features': 768, 'out_features': 512, 'bias': True, 'activation_function': 'torch.nn.modules.activation.Tanh'})
)

📄 Citing & Authors