Bharat-Tiny-LLM-v2/README.md

---
tags:
- hindi
- devanagari
- brahmi
- qwen2.5
- transformers
- safetensors
- edge-ai
- india
- llm
- pytorch
license: apache-2.0
language:
- hi
- en
pipeline_tag: text-generation
library_name: transformers
---

<p align="center">
  <img src="https://img.shields.io/badge/Model_Size-1.5B-blue?style=flat-square">
  <img src="https://img.shields.io/badge/Format-PyTorch_Safetensors-brightgreen?style=flat-square">
  <img src="https://img.shields.io/badge/Hindi_Compression-33.8%25-orange?style=flat-square">
  <img src="https://img.shields.io/badge/License-Apache_2.0-success?style=flat-square">
  <img src="https://img.shields.io/badge/Made_by-eulogik-purple?style=flat-square">
</p>

<h1 align="center">🇮🇳 Bharat-Tiny-LLM v2 (PyTorch / Transformers Base)</h1>
<p align="center"><em>1.5B Parameter Open-Weights Indic LLM featuring Brahmi Token Injection for 33.8% token compression and +36% faster Hindi inference.</em></p>

---

![Bharat-Tiny-LLM v2 Technical Infographic](bharat_tiny_llm_v2_tech_infographic.jpg)

---

## 🎯 Model Overview

**Bharat-Tiny-LLM v2** is an open-weights 1.5B parameter language model built on top of Qwen2.5-1.5B, optimized specifically for Hindi and Hinglish text generation.

By introducing **Brahmi Token Injection** — a technique that surgically injects 300 Devanagari subword tokens into the model's vocabulary — Bharat-Tiny-LLM v2 eliminates the severe "Token Tax" imposed by standard English-centric tokenizers on Indian scripts.

This repository contains the **unquantized PyTorch / HuggingFace Transformers open weights**, compatible with Linux, Windows, CUDA GPUs, vLLM, TGI, and Google Colab.

---

## 🎨 Architectural Pipeline & Training Methodology

![Brahmi Token Injection Architecture](bharat_tiny_llm_v2_pipeline_diagram.jpg)

### Interactive Training Pipeline Flowchart

```mermaid
flowchart LR
    A["Raw Hindi Corpus"] --> B["Brahmi Subword Mining<br>(Top 300 Devanagari Tokens)"]
    B --> C["Tokenizer Vocabulary Expansion<br>(151,936 ➔ 152,236)"]
    C --> D["Stage 1: Embedding Alignment<br>(Freeze Backbone, Train 300 Embeddings)"]
    D --> E["Stage 2: LoRA Fine-Tuning<br>(Rank=16 on Attention q,k,v,o proj)"]
    E --> F["Fused PyTorch Base Weights<br>(eulogik/Bharat-Tiny-LLM-v2)"]
    F --> G["Q4 Affine Quantization<br>(eulogik/Bharat-Tiny-LLM-v2-MLX)"]
```

---

## ✨ Key Benchmarks & Technical Advantages

| Metric | Base Qwen2.5-1.5B | Bharat-Tiny-LLM v2 | Technical Advantage |
|:---|:---|:---|:---|
| **Tokens for 1,000 Hindi Chars** | ~950 tokens | **~630 tokens** | **33.8% Fewer Tokens (up to 58% on chat prompts)** |
| **Inference Throughput (Hindi)** | 50 tok/s | **68 tok/s** | **+36% Speed Boost** |
| **Validation Loss (Hindi Corpus)** | 2.776 | **1.837** | **52.5% Loss Reduction (Perplexity: 16.1 → 6.3)** |
| **Hardware Compatibility** | CUDA / CPU / MPS | CUDA / CPU / MPS | **Universal PyTorch / vLLM / GGUF support** |

---

## 🚀 Quick Start with PyTorch & Transformers

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "eulogik/Bharat-Tiny-LLM-v2"

# Load Tokenizer & Model Weights
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Generate Hindi Text
prompt = "भारत की सांस्कृतिक विविधता के बारे में बताइए:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=150,
    temperature=0.7,
    do_sample=True
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

---

## 📊 Token Compression Benchmarks

| Prompt (Hindi / Hinglish) | Base Qwen Tokens | Bharat-v2 Tokens | Savings |
|:---|:---|:---|:---|
| `"ज़रूरी बात है क्या करते हो"` | 26 tokens | **11 tokens** | **58% Savings** |
| `"नमस्ते, आप कैसे हैं?"` | 15 tokens | **7 tokens** | **53% Savings** |
| `"भारत की राजधानी नई दिल्ली है"` | 22 tokens | **14 tokens** | **36% Savings** |
| `"bhai aaj ka weather kaisa hai?"` | 12 tokens | **8 tokens** | **33% Savings** |

---

## 🔗 Model Family Repositories

- **Apple Silicon MLX Quantized (880MB)**: [`eulogik/Bharat-Tiny-LLM-v2-MLX`](https://huggingface.co/eulogik/Bharat-Tiny-LLM-v2-MLX)
- **GGUF / llama.cpp**: [`eulogik/Bharat-Tiny-LLM-GGUF`](https://huggingface.co/eulogik/Bharat-Tiny-LLM-GGUF)

---

## 📜 License & Citation

Licensed under **Apache 2.0**. Free for commercial, enterprise, and research use.

```bibtex
@misc{kishore2026brahmi,
    title={Brahmi: Efficient Devanagari Token Injection for Multilingual LLMs},
    author={Gautam Kishore},
    year={2026},
    publisher={eulogik},
    howpublished={\url{https://huggingface.co/eulogik/Bharat-Tiny-LLM-v2}}
}
```