Ruri V3 70m

🚀 Ruri: Japanese General Text Embeddings

Ruri v3 is a general - purpose Japanese text embedding model built on top of ModernBERT - Ja. It offers a solution for Japanese text embedding tasks, providing state - of - the - art performance and several technical advantages.

🚀 Quick Start

You can use our models directly with the transformers library v4.48.0 or higher:

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

Additionally, if your GPUs support Flash Attention 2, we recommend using our models with Flash Attention 2.

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-v3-70m", 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, 384]

similarities = F.cosine_similarity(embeddings.unsqueeze(0), embeddings.unsqueeze(1), dim=2)
print(similarities)
# [[1.0000, 0.9555, 0.8430, 0.6772, 0.7077],
#  [0.9555, 1.0000, 0.8333, 0.6636, 0.6971],
#  [0.8430, 0.8333, 1.0000, 0.8554, 0.8639],
#  [0.6772, 0.6636, 0.8554, 1.0000, 0.9500],
#  [0.7077, 0.6971, 0.8639, 0.9500, 1.0000]]

✨ Features

• State - of - the - art performance: Offers top - notch performance for Japanese text embedding tasks.
• Longer sequence support: Supports sequence lengths up to 8192 tokens, compared to the previous versions (v1, v2) which were limited to 512.
• Expanded vocabulary: Has an expanded vocabulary of 100K tokens, as opposed to 32K in v1 and v2, making input sequences shorter and improving efficiency.
• Integrated FlashAttention: Following ModernBERT's architecture, it enables faster inference and fine - tuning.
• Simplified tokenizer: Uses a tokenizer based solely on SentencePiece. Unlike previous versions that relied on Japanese - specific BERT tokenizers and required pre - tokenized input, Ruri v3 performs tokenization with SentencePiece only, eliminating the need for an external word segmentation tool.

📦 Installation

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

If your GPUs support Flash Attention 2:

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

📚 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': 384, '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})
)

🔧 Technical Details

Ruri v3 is built on the foundation of ModernBERT - Ja. It has made several improvements in terms of sequence length support, vocabulary size, and tokenization method. The integration of FlashAttention also significantly enhances the efficiency of inference and fine - tuning.

📄 License

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

📚 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}, 
}