Ruri V3 310m

🚀 Ruri: Japanese General Text Embeddings

Ruri v3 is a general-purpose Japanese text embedding model built on top of ModernBERT-Ja. It offers state-of-the-art performance for Japanese text embedding tasks, supports sequence lengths up to 8192 tokens, and has an expanded vocabulary of 100K tokens.

🚀 Quick Start

You can use our models directly with the transformers library v4.48.0 or higher. First, install the necessary libraries:

pip install -U "transformers>=4.48.0" sentence-transformers

If your GPUs support Flash Attention 2, we recommend using our models with it:

pip install flash-attn --no-build-isolation

Then you can load this model and run inference:

import torch
import torch.nn.functional as F
from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
device = "cuda" if torch.cuda.is_available() else "cpu"
model = SentenceTransformer("cl-nagoya/ruri-v3-310m", device=device)

# Ruri v3 employs a 1+3 prefix scheme to distinguish between different types of text inputs:
# "" (empty string) is used for encoding semantic meaning.
# "トピック: " is used for classification, clustering, and encoding topical information.
# "検索クエリ: " is used for queries in retrieval tasks.
# "検索文書: " is used for documents to be retrieved.
sentences = [
    "川べりでサーフボードを持った人たちがいます",
    "サーファーたちが川べりに立っています",
    "トピック: 瑠璃色のサーファー",
    "検索クエリ: 瑠璃色はどんな色？",
    "検索文書: 瑠璃色（るりいろ）は、紫みを帯びた濃い青。名は、半貴石の瑠璃（ラピスラズリ、英: lapis lazuli）による。JIS慣用色名では「こい紫みの青」（略号 dp-pB）と定義している[1][2]。",
]

embeddings = model.encode(sentences, convert_to_tensor=True)
print(embeddings.size())
# [5, 768]

similarities = F.cosine_similarity(embeddings.unsqueeze(0), embeddings.unsqueeze(1), dim=2)
print(similarities)
# [[1.0000, 0.9603, 0.8157, 0.7074, 0.6916],
#  [0.9603, 1.0000, 0.8192, 0.7014, 0.6819],
#  [0.8157, 0.8192, 1.0000, 0.8701, 0.8470],
#  [0.7074, 0.7014, 0.8701, 1.0000, 0.9746],
#  [0.6916, 0.6819, 0.8470, 0.9746, 1.0000]]

✨ Features

• State-of-the-art performance for Japanese text embedding tasks.
• Supports sequence lengths up to 8192 tokens
  • Previous versions of Ruri (v1, v2) were limited to 512.
• Expanded vocabulary of 100K tokens, compared to 32K in v1 and v2
  • The larger vocabulary make input sequences shorter, improving efficiency.
• Integrated FlashAttention, following ModernBERT's architecture
  • Enables faster inference and fine-tuning.
• Tokenizer based solely on SentencePiece
  • Unlike previous versions, which relied on Japanese-specific BERT tokenizers and required pre-tokenized input, Ruri v3 performs tokenization with SentencePiece only—no external word segmentation tool is required.

📦 Installation

pip install -U "transformers>=4.48.0" sentence-transformers

If your GPUs support Flash Attention 2:

pip install flash-attn --no-build-isolation

💻 Usage Examples

Basic Usage

import torch
import torch.nn.functional as F
from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
device = "cuda" if torch.cuda.is_available() else "cpu"
model = SentenceTransformer("cl-nagoya/ruri-v3-310m", device=device)

# Ruri v3 employs a 1+3 prefix scheme to distinguish between different types of text inputs:
# "" (empty string) is used for encoding semantic meaning.
# "トピック: " is used for classification, clustering, and encoding topical information.
# "検索クエリ: " is used for queries in retrieval tasks.
# "検索文書: " is used for documents to be retrieved.
sentences = [
    "川べりでサーフボードを持った人たちがいます",
    "サーファーたちが川べりに立っています",
    "トピック: 瑠璃色のサーファー",
    "検索クエリ: 瑠璃色はどんな色？",
    "検索文書: 瑠璃色（るりいろ）は、紫みを帯びた濃い青。名は、半貴石の瑠璃（ラピスラズリ、英: lapis lazuli）による。JIS慣用色名では「こい紫みの青」（略号 dp-pB）と定義している[1][2]。",
]

embeddings = model.encode(sentences, convert_to_tensor=True)
print(embeddings.size())
# [5, 768]

similarities = F.cosine_similarity(embeddings.unsqueeze(0), embeddings.unsqueeze(1), dim=2)
print(similarities)
# [[1.0000, 0.9603, 0.8157, 0.7074, 0.6916],
#  [0.9603, 1.0000, 0.8192, 0.7014, 0.6819],
#  [0.8157, 0.8192, 1.0000, 0.8701, 0.8470],
#  [0.7074, 0.7014, 0.8701, 1.0000, 0.9746],
#  [0.6916, 0.6819, 0.8470, 0.9746, 1.0000]]

📚 Documentation

Model Series

We provide Ruri-v3 in several model sizes. Below is a summary of each model.

Benchmarks

JMTEB

Evaluated with JMTEB.

Model Details

Model Description

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 8192, 'do_lower_case': False}) with Transformer model: ModernBertModel 
  (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, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
)

📄 License

This model is published under the Apache License, Version 2.0.

🔧 Technical Details

Ruri v3 offers several key technical advantages:

• State-of-the-art performance for Japanese text embedding tasks.
• Supports sequence lengths up to 8192 tokens
  • Previous versions of Ruri (v1, v2) were limited to 512.
• Expanded vocabulary of 100K tokens, compared to 32K in v1 and v2
  • The larger vocabulary make input sequences shorter, improving efficiency.
• Integrated FlashAttention, following ModernBERT's architecture
  • Enables faster inference and fine-tuning.
• Tokenizer based solely on SentencePiece
  • Unlike previous versions, which relied on Japanese-specific BERT tokenizers and required pre-tokenized input, Ruri v3 performs tokenization with SentencePiece only—no external word segmentation tool is required.

📖 Citation

@misc{
  Ruri,
  title={{Ruri: Japanese General Text Embeddings}}, 
  author={Hayato Tsukagoshi and Ryohei Sasano},
  year={2024},
  eprint={2409.07737},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2409.07737}, 
}