---
license: llama3.2
base_model: meta-llama/Llama-3.2-3B-Instruct
language:
- en
- nng
pipeline_tag: text-generation
tags:
- llama-3
- llama-3.2
- fine-tuned
- maringgpt
- maring-language
- nlp
---

<div align="center">

# 🌌 MaringGPT

**A modern, lightweight AI model built to preserve, translate, and advance digital support for the Maring language.**

[![Base Model](https://img.shields.io/badge/Base%20Model-Llama--3.2--3B--Instruct-blue?style=for-the-badge&logo=meta)](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct)
[![License](https://img.shields.io/badge/License-Llama%203.2-green?style=for-the-badge)](https://llama.meta.com/license)
[![Framework](https://img.shields.io/badge/Framework-Unsloth%20%2F%20TRL-orange?style=for-the-badge)](https://github.com/unslothai/unsloth)

</div>

---

## 📜 About the Maring Language & History

> *"Meiringba" / "Maringa" — People who keep the unquenched fire alive.*

**Maring** (ISO 639-3: `nng`) is a Sino-Tibetan language spoken primarily by the Maring community in the Tengnoupal and Chandel districts of Manipur, Northeast India.

* **Ethno-Linguistic Roots:** Linguistically classified within the Tibeto-Burman branch, Maring shares unique connections with both Tangkhulic and Kuki-Chin languages.
* **Cultural Identity:** Traditionally an oral-rich language, its speakers are deeply connected to the hills of Southeast Manipur.
* **The Digital Gap:** Like many indigenous languages of Northeast India, Maring remains under-represented in modern computational linguistics and digital datasets. **MaringGPT** is created to help bridge this gap by bringing native context and modern AI capabilities into a fine-tuned LLM.

---

## 📸 Demo Preview

| Preview 01 | Preview 02 |
| :---: | :---: |
| ![MaringGPT Demo 1](./01.png) | ![MaringGPT Demo 2](./02.png) |

---

## ✨ Key Features

* 🗣️ **Language Awareness:** Fine-tuned to understand context and structured text associated with Maring language instruction.
* ⚡ **Ultra-Efficient (3B Parameters):** Designed on Meta's Llama 3.2 3B architecture, enabling low latency and local execution on consumer GPUs or mobile platforms.
* 🎯 **Instruction-Tuned:** Structured dialogue handling via ChatML / Llama-3 system prompts for crisp, natural conversations.

---

## ⚙️ Training Setup

The model was fine-tuned using parameter-efficient optimization techniques:

| Parameter | Specification |
| :--- | :--- |
| **Base Model** | `meta-llama/Llama-3.2-3B-Instruct` |
| **Method** | QLoRA (4-bit Quantization) via Unsloth |
| **Precision** | `bfloat16` |
| **Data Format** | Custom conversational instruction-response pairs |

---

## 🚀 Quickstart

Run **MaringGPT** locally using Python and the `transformers` library:

### 1. Installation

```bash
pip install --upgrade transformers torch accelerate