hieptran318204/vti-embedding-bge-v6

Model Card: vti-bge-embedding-v6

vti-bge-embedding-v6 is an embedding model fine-tuned from the Vietnamese_Embedding_v2 model (https://huggingface.co/BAAI/bge-reranker-v2-m3) to enhance retrieval capabilities for Vietnamese in an especially domain (VTI's environment).

• The original model was trained on approximately 1,100,000 triplets of queries, positive documents, and negative documents for Vietnamese.
• The model was trained with a maximum sequence length of 2304 (256 for query and 2048 for passages).
• The new model was fine-tuned with about 6,000 triplets of queries, positive documents, and degative documents in an tecnology company.

Model Details

Model Description

• Model Type: Sentence Transformer
• Base model: BAAI/bge-m3
• Maximum Sequence Length: 2048 tokens
• Output Dimensionality: 1024 dimensions
• Similarity Function: Dot product Similarity
• Language: Vietnamese
• Licence: Apache 2.0

Usage

from sentence_transformers import SentenceTransformer
import torch
model = SentenceTransformer("hieptran318204/vti-bge-embedding-v6")
model.max_seq_length = 2048
sentences_1 = ["Trí tuệ nhân tạo là gì", "Lợi ích của giấc ngủ"]
sentences_2 = ["Trí tuệ nhân tạo là công nghệ giúp máy móc suy nghĩ và học hỏi như con người. Nó hoạt động bằng cách thu thập dữ liệu, nhận diện mẫu và đưa ra quyết định.", 
               "Giấc ngủ giúp cơ thể và não bộ nghỉ ngơi, hồi phục năng lượng và cải thiện trí nhớ. Ngủ đủ giấc giúp tinh thần tỉnh táo và làm việc hiệu quả hơn."]
query_embedding = model.encode(sentences_1)
doc_embeddings = model.encode(sentences_2)
similarity = query_embedding @ doc_embeddings.T
print(similarity)
'''
array([[0.66212064, 0.33066642],
       [0.25866613, 0.5865289 ]], dtype=float32)
'''

Evaluation:

• Dataset: Entire training dataset of Legal Zalo 2021. Our model was not trained on this dataset.

Vietnamese_Reranker and Vietnamese_Embedding_v2 was trained on 1,100,000 triplets. Although the score on the legal domain drops a bit on Vietnamese_Embedding (Phase 2), since this phase data is much larger, it is good for other domains.

You can reproduce the evaluation result by running code python evaluation_model.py (data downloaded from Kaggle).

Contact

Email: hiep.tranvan@vti.com.vn

Citation

@misc{vti-bge-embedding,
  title={vti-bge-embedding-v6: Embedding model in Vietnamese language.},
  author={Nguyen Nho Trung, Nguyen Nhat Quang, Nguyễn Văn Huy},
  year={2025},
  publisher={Huggingface},
}