README.md

---
language: hi
license: mit
tags:
  - hindi
  - embeddings
  - sentence-embeddings
  - semantic-search
  - text-similarity
datasets:
  - custom
pipeline_tag: sentence-similarity
library_name: transformers
---

# Hindi Sentence Embeddings Model

This is a custom state-of-the-art sentence embedding model trained specifically for Hindi text. It leverages an advanced transformer architecture with specialized pooling strategies to create high-quality semantic representations of Hindi sentences.

## Features

- Specialized for Hindi language text
- Advanced transformer architecture with optimized attention mechanism
- Multiple pooling strategies for enhanced semantic representations
- Creates normalized vector representations for semantic similarity
- Supports semantic search and text similarity applications

## Usage

### Installation

```bash
pip install torch sentencepiece scikit-learn matplotlib
git lfs install 
git clone https://huggingface.co/DeepMostInnovations/hindi-embedding-foundational-model-10B
cd hindi-embedding-foundational-model-10B
```

### Enhanced RAG System

This model now includes an enhanced RAG (Retrieval Augmented Generation) system that integrates Unsloth's optimized Llama-3.2-1B-Instruct model for question answering on top of Hindi document retrieval.

#### Setup and Installation

1. Install additional dependencies:
```bash
pip install unsloth transformers bitsandbytes accelerate langchain langchain-community faiss-cpu
```

2. Index your documents:
```bash
python hindi-rag-system.py --model_dir /path/to/your/model --tokenizer_dir /path/to/tokenizer --data_dir ./data --output_dir ./output --index
```

3. Run in QA mode with LLM:
```bash
python hindi-rag-system.py --model_dir /path/to/your/model --tokenizer_dir /path/to/tokenizer --output_dir ./output --interactive --qa
```

### Basic Embedding Usage

```python
from hindi_embeddings import HindiEmbedder

# Initialize the embedder
model = HindiEmbedder("path/to/hindi-embedding-foundational-model-10B")

# Encode sentences to embeddings
sentences = [
    "मुझे हिंदी भाषा बहुत पसंद है।",
    "मैं हिंदी भाषा सीख रहा हूँ।"
]
embeddings = model.encode(sentences)
print(f"Embedding shape: {embeddings.shape}")

# Compute similarity between sentences
similarity = model.compute_similarity(sentences[0], sentences[1])
print(f"Similarity: {similarity:.4f}")

# Perform semantic search
query = "भारत की राजधानी"
documents = [
    "दिल्ली भारत की राजधानी है।",
    "मुंबई भारत का सबसे बड़ा शहर है।",
    "हिमालय पर्वत भारत के उत्तर में स्थित है।"
]
results = model.search(query, documents)
for i, result in enumerate(results):
    print(f"{i+1}. Score: {result['score']:.4f}")
    print(f"   Document: {result['document']}")

# Visualize embeddings
example_sentences = [
    "मुझे हिंदी में पढ़ना बहुत पसंद है।",
    "आज मौसम बहुत अच्छा है।",
    "भारत एक विशाल देश है।"
]
model.visualize_embeddings(example_sentences)
```

## Model Details

This model uses an advanced transformer-based architecture with the following enhancements:

- Pre-layer normalization for stable training
- Specialized attention mechanism with relative positional encoding
- Multiple pooling strategies (weighted, mean, attention-based)
- L2-normalized vectors for cosine similarity

Technical specifications:
- Embedding dimension: 768
- Hidden dimension: 768
- Layers: 12
- Attention heads: 12
- Vocabulary size: 50,000
- Context length: 128 tokens

## Applications

- Semantic search and information retrieval
- Text clustering and categorization
- Recommendation systems
- Question answering
- Document similarity comparison
- Content-based filtering
- RAG systems for Hindi language content

## License

This model is released under the MIT License.

## Citation

If you use this model in your research or application, please cite us:

```
@misc{DeepMostInnovations2025hindi,
  author = {DeepMost Innovations},
  title = {Hindi Sentence Embeddings Model},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/DeepMostInnovations/hindi-embedding-foundational-model-10B}}
}
```