nmixx-bge-m3

NMIXX-BGE-M3

This repository contains a bge-m3-based SentenceTransformer model fine-tuned with a triplet-loss setup on the nmixx-fin/NMIXX_train dataset. It produces high-quality sentence embeddings for Korean financial text with multilingual capabilities, optimized for semantic similarity tasks in the finance domain.

---

How to Use

from transformers import AutoTokenizer, AutoModel
import torch
import torch.nn.functional as F

# 1. Load tokenizer & model from Hugging Face Hub
repo_name = "nmixx-fin/nmixx-bge-m3"
tokenizer = AutoTokenizer.from_pretrained(repo_name)
model = AutoModel.from_pretrained(repo_name)

# 2. Move to GPU if available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
model.eval()

# 3. Prepare input sentences
sentences = [
    "이 모델은 한국 금융 도메인에 특화된 다국어 임베딩을 제공합니다.",
    "NMIXX 데이터셋으로 fine-tuning된 multilingual sentence transformer입니다."
]

# 4. Tokenize
encoded_input = tokenizer(
    sentences,
    padding=True,
    truncation=True,
    max_length=8192,  # BGE-M3 supports longer sequences
    return_tensors="pt"
)
input_ids = encoded_input["input_ids"].to(device)
attention_mask = encoded_input["attention_mask"].to(device)

# 5. Forward pass (token embeddings)
with torch.no_grad():
    model_output = model(input_ids=input_ids, attention_mask=attention_mask)

# 6. CLS Pooling (BGE models use CLS token)
sentence_embeddings = model_output[0][:, 0]  # Use CLS token (first token)

# 7. L2 Normalization
sentence_embeddings = F.normalize(sentence_embeddings, p=2, dim=1)

print("Sentence embeddings shape:", sentence_embeddings.shape)
print(sentence_embeddings.cpu())

Features

• Multilingual Support: Based on BGE-M3, supporting Korean and English
• Extended Sequence Length: Supports up to 8192 tokens
• Korean Financial Domain: Fine-tuned specifically for Korean financial text
• High Performance: Optimized for semantic similarity tasks in finance

Model Details

• Base Model: BAAI/bge-m3
• Fine-tuning Dataset: nmixx-fin/NMIXX_train
• Training Method: Triplet loss with hard negative mining
• Embedding Dimension: 1024
• Max Sequence Length: 8192 tokens (Stable performance can only be guaranteed up to 2048 tokens)

⚠️ Developer's Note: 이거 금융 STS 말고 잘하는 거 없습니다.

@article{lee2025nmixxdomainadaptedneuralembeddings,
  title        = {NMIXX: Domain-Adapted Neural Embeddings for Cross-Lingual eXploration of Finance},
  author       = {Hanwool Lee and Sara Yu and Yewon Hwang and Jonghyun Choi and Heejae Ahn and Sungbum Jung and Youngjae Yu},
  year         = {2025},
  eprint       = {2507.09601},
  archivePrefix= {arXiv},
  primaryClass = {cs.CL},
  url          = {https://arxiv.org/abs/2507.09601},
}