AhinsaAI/Ahinsa-nli-triplet

Great! Based on your clarification, here is an updated and specific README.md for your project focused on fine-tuning sentence-transformers/multilingual-mpnet-base-v2 for Hindi using triplet loss:

---

🇮🇳 Hindi Semantic Similarity Dataset for Multilingual MPNet Fine-tuning

🧠 Project Overview

This project focuses on fine-tuning the sentence-transformers/multilingual-mpnet-base-v2 model using Hindi-language sentence triplets. The objective is to create a robust Hindi sentence embedding model capable of semantic similarity, clustering, and retrieval tasks.

---

🗂 Dataset Structure

The dataset is structured as triplets suitable for triplet loss:

Column	Description
anchor	A Hindi sentence (reference)
positive	A semantically similar sentence to the anchor
negative	A semantically different or contradictory sentence to anchor

🔍 Sample:

anchor,positive,negative
"एक बुजुर्ग आदमी बर्गर तल रहा है।","एक आदमी ग्रिल पर बर्गर बना रहा है।","एक बूढ़ा आदमी कच्चा बर्गर खा रहा है।"

---

🧪 Data Splits

The cleaned dataset is split into:

• export/train.csv – 80%
• export/dev.csv – 10%
• export/test.csv – 10%

Each file follows the same structure (anchor, positive, negative).

---

🤖 Model & Training

🏷 Base Model:

• sentence-transformers/multilingual-mpnet-base-v2
• Supports 50+ languages, including Hindi

🧪 Objective:

• Triplet Loss fine-tuning using sentence-transformers framework
• Input triplets: (anchor, positive, negative)
• Output: embeddings that place anchor closer to positive than to negative

🛠 Training Framework:

• Sentence-Transformers
• Triplet loss with a margin (default: 0.5)
• Evaluation with cosine similarity and embedding ranking

---

📦 How to Load

You can load the dataset with Hugging Face:

from datasets import load_dataset

dataset = load_dataset('csv', data_files={
    'train': 'export/train.csv',
    'validation': 'export/dev.csv',
    'test': 'export/test.csv'
})

Or directly with pandas:

import pandas as pd
train = pd.read_csv("export/train.csv")

---

💡 Use Cases

• Hindi semantic search
• Paraphrase mining and deduplication
• Hindi text clustering
• Text-based retrieval systems

---

⚠️ Limitations

• Relatively small dataset size; more data will improve performance.
• Sentence triplets are heuristic or manually generated, not crowdsourced.
• May not cover complex linguistic phenomena or dialectal variations.

---

📚 Citation

@dataset{,
  title     = {Hindi Triplet Dataset for Multilingual MPNet Fine-tuning},
  author    = {Your Name},
  year      = {2025},
  url       = {https://github.com/yourusername/hindi-triplet-mpnet}
}

---

📬 Contact

For questions, contributions, or feedback, contact: 📧 [email protected] 🐙 GitHub: @yourusername

---