---
license: apache-2.0
base_model: unsloth/Qwen3-1.7B
language:
- hi
- en
- bn
- ta
- te
- mr
- gu
pipeline_tag: text-generation
tags:
- indian-languages
- hinglish
- reasoning
- gguf
- quantization
- unsloth
- legal
- agriculture
---

# Indica-1.7B-GGUF 🇮🇳

**Indica-1.7B** is a lightweight, high-performance SLM (Small Language Model) specifically post-trained for the Indian context. Built upon the **Qwen3-1.7B** dense transformer architecture, this model has undergone a rigorous multi-stage alignment (fine-tuning for behavior and preference) process to excel in Hindi, Hinglish, and various regional dialects while maintaining strong CoT (Chain-of-Thought) reasoning capabilities.

This repository provides the model in **GGUF** format, optimized for local inference (the process of generating text) on consumer hardware using tools like **Ollama**, **llama.cpp**, and **LM Studio**.

## 🚀 Model Details
- **Architecture:** 1.7 Billion parameters (the internal variables a neural network learns), utilizing a dense causal transformer design.
- **Specialized Domains:** Tailored for Indian Law (IPC/BNS), Agriculture (MSP/PM-Kisan), and National Examinations (UPSC/JEE).
- **Multilingual Mastery:** Fluent in Hindi-English code-switching (Hinglish) and supports multiple regional Indian languages.
- **Thinking Paradigm:** Utilizes a native "thinking mode" via `<think>` tags for complex reasoning tasks before outputting a final answer.
- **Context Length:** Supports up to 2048 tokens natively.
- **Developer:** Prashant (prash616).

## 🛠 Training Pipeline
The model was developed through a specialized four-stage alignment strategy:

1. **Foundational Pre-training:** Fine-tuned on Hindi Wikipedia (`wikimedia/wikipedia`) to establish deep linguistic roots and vocabulary density.
2. **Supervised Fine-Tuning (SFT):** Trained on high-quality instruction datasets (`ai4bharat/indic-instruct-data-v0.1`, `viber1/indian-law-dataset`) covering Indian law, agriculture, and everyday Hinglish conversations.
3. **GRPO (Group Relative Policy Optimization):** Aligned using Reinforcement Learning to reward logical reasoning and the use of internal thinking tags using GSM8K datasets.
4. **DPO (Direct Preference Optimization):** Final behavioral polish using `HuggingFaceH4/ultrafeedback_binarized` to ensure a helpful, polite, and culturally aware persona.

## 📦 Quantization Details
These GGUF files were created using `llama.cpp` through the Unsloth library. Quantization (the process of reducing the precision of the model's numbers) allows the model to run on machines with limited VRAM (Video Random Access Memory).

| Filename | Bit-Size | File Size | Use Case |
| :--- | :--- | :--- | :--- |
| `Indica-1.7B-Q4_K_M.gguf` | 4-bit | ~1.1 GB | **Recommended.** Balanced quality and extreme speed. Ideal for standard laptops, MacBooks, and low-RAM devices. |
| `Indica-1.7B-Q8_0.gguf` | 8-bit | ~1.8 GB | **Maximum Quality.** Retains near-perfect precision from the 16-bit model. Recommended for technical legal or mathematical queries. |

## 💻 How to Use

### 1. With Ollama (Easiest)
Ensure you have [Ollama](https://ollama.com/) installed, then run:
```bash
ollama run hf.co/prash616/Indica-1.7B-GGUF




## 2. With LM Studio
Download LM Studio.
Search for prash616/Indica-1.7B-GGUF in the search bar.
Download the Q4_K_M file and load it into the local server.

##3. Chat Template (For Developers)
If you are writing custom Python inference scripts, the model uses the standard qwen-3 chat template. Ensure your system prompt is set correctly: