Tooka-SBERT-V2-Small

This model is a Sentence Transformers model trained for semantic textual similarity and embedding tasks. It maps sentences and paragraphs to a dense vector space, where semantically similar texts are close together.

The model is trained in two sizes: Small and Large

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install sentence-transformers==3.4.1

Then you can load this model and run inference.

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("PartAI/Tooka-SBERT-V2-Small")
# Run inference
sentences = [
    'درنا از پرندگان مهاجر با پاهای بلند و گردن دراز است.',
    'درناها با قامتی بلند و بال‌های پهن، از زیباترین پرندگان مهاجر به شمار می‌روند.',
    'درناها پرندگانی کوچک با پاهای کوتاه هستند که مهاجرت نمی‌کنند.'
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]

🛠️ Training Details

The training is performed in two stages:

1. Pretraining on the Targoman News dataset
2. Fine-tuning on multiple synthetic datasets

Stage 1: Pretraining

• We use an asymmetric setup.
• Input formatting:
  • Titles are prepended with "سوال: "
  • Texts are prepended with "متن: "
• Loss function: CachedMultipleNegativesRankingLoss

Stage 2: Fine-tuning

• Loss functions:
  • CachedMultipleNegativesRankingLoss
  • CoSENTLoss
• Used across multiple synthetic datasets

📊 Evaluation

We evaluate our model on the PTEB Benchmark. Our model outperforms mE5-Base on average across PTEB tasks.

For Retrieval and Reranking tasks, we follow the same asymmetric structure, prepending:

• "سوال: " to queries
• "متن: " to documents

Model	#Params	Pair-Classification-Avg	Classification-Avg	Retrieval-Avg	Reranking-Avg	CrossTasks-Avg
Tooka-SBERT-V2-Large	353M	80.24	74.73	59.80	73.44	72.05
Tooka-SBERT-V2-Small	123M	75.69	72.16	61.24	73.40	70.62
jina-embeddings-v3	572M	71.88	79.27	65.18	64.62	70.24
multilingual-e5-base	278M	70.76	69.71	63.90	76.01	70.09
Tooka-SBERT-V1-Large	353M	81.52	71.54	45.61	60.44	64.78

Task-Specific Datasets in PTEB

• Pair-Classification:

  • FarsTail

• Classification:

  • MassiveIntentClassification
  • MassiveScenarioClassification
  • MultilingualSentimentClassification
  • PersianFoodSentimentClassification

• Retrieval:

  • MIRACLRetrieval
  • NeuCLIR2023Retrieval
  • WikipediaRetrievalMultilingual

• Reranking:

  • MIRACLReranking
  • WikipediaRerankingMultilingual

Citation

BibTeX

Sentence Transformers

@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",
}

CachedMultipleNegativesRankingLoss

@misc{gao2021scaling,
    title={Scaling Deep Contrastive Learning Batch Size under Memory Limited Setup}, 
    author={Luyu Gao and Yunyi Zhang and Jiawei Han and Jamie Callan},
    year={2021},
    eprint={2101.06983},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}