xmanii/maux-gte-persian-v3-fp16

maux-gte-persian-v3 (fp16)

A high-performance Persian sentence embedding model based on Alibaba-NLP/gte-multilingual-base, released in fp16 for efficient inference.

---

Model Overview

This is the fp16 (half-precision) version of maux-gte-persian-v3, a Sentence Transformers model fine-tuned from Alibaba-NLP/gte-multilingual-base for robust Persian sentence and paragraph embeddings.
The fp16 format enables faster and more memory-efficient inference, especially on modern GPUs.

Key Features:

• Base Model: Alibaba-NLP/gte-multilingual-base
• Fine-tuned on: mshojaei77/Persian_sft (80,000 Persian sentence pairs)
• Output Dimension: 768
• Max Sequence Length: 8192 tokens
• Similarity Function: Cosine Similarity
• Loss Function: MultipleNegativesRankingLoss
• Format: fp16 (model.safetensors)

---

Performance

• Excellent performance on Persian semantic similarity, search, and clustering tasks.
• Outperforms or matches jinaai-v3 in most Persian benchmarks (see comparison charts).
• Efficient for large-scale inference due to fp16 weights.

---

Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 8192, 'do_lower_case': False}) with Transformer model: NewModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, ...})
  (2): Normalize()
)

---

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("xmanii/maux-gte-persian-v3-fp16", trust_remote_code=True)
sentences = [
    'برج میلاد در تهران هست',
    'یکی از برج های مسکونی تهران برج تهران است',
    'تهران برج های زیادی دارد'
]
embeddings = model.encode(sentences)
print(embeddings.shape)  # [3, 768]

# Compute cosine similarity
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)  # [3, 3]

---

Training Details

• Dataset: mshojaei77/Persian_sft
• Loss: MultipleNegativesRankingLoss (scale=20.0, similarity_fct="cos_sim")
• Batch size: 64
• Precision: bf16 during training, fp16 for this release
• Frameworks: Python 3.10, Sentence Transformers 4.1.0, Transformers 4.51.3, PyTorch 2.7.0+cu126

---

Files

• model.safetensors (fp16 weights)
• All necessary config and tokenizer files
• Custom code: modeling.py, configuration.py (required for loading)

---

Citation

If you use this model, please cite:

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}
@misc{henderson2017efficient,
    title={Efficient Natural Language Response Suggestion for Smart Reply},
    author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil},
    year={2017},
    eprint={1705.00652},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

---

Acknowledgements

• Special thanks to mshojaei77 for the Persian_sft dataset.
• Built on top of Alibaba-NLP/gte-multilingual-base.

---

License

This model is distributed under the same license as the base model and dataset.

---

For questions or feedback, please open an issue or discussion on the Hugging Face model page.