---
license: apache-2.0
base_model: unsloth/Qwen3-1.7B
language:
- hi
- en
pipeline_tag: text-generation
tags:
- indian-languages
- hinglish
- reasoning
- experimental
- research
- unsloth
---

# 🧪 Indica-1.7B: An Experimental Research Model 🇮🇳

> **NOTICE: This is an experimental model released for research and development purposes. It serves as a proof-of-concept for a 4-stage post-training pipeline on Small Language Models (SLMs).**

**Indica-1.7B** is a lightweight model developed by **Prashant** to explore the limits of persona-injection and cultural alignment in ultra-small parameter architectures (1.7B). Built on **Qwen3-1.7B**, this model was subjected to a rigorous training regime including **SFT**, **GRPO (Reasoning)**, and **DPO (Alignment)**.

---

## 🔬 The Research Objective
The goal of this project was to test whether a 1.7B model could successfully balance three conflicting objectives:
1. **Domain Expertise:** Knowledge of Indian Law (IPC/BNS) and Agriculture.
2. **Linguistic Persona:** Natural, colloquial Hinglish/Hindi code-switching.
3. **Logic & Reasoning:** Utilizing a native "Thinking" trace via Reinforcement Learning.

## 🛠️ Post-Training Pipeline
The model underwent a specialized four-stage alignment strategy:
* **Stage 1: SFT (Knowledge):** Trained on Indian Law and Agriculture datasets.
* **Stage 2: GRPO (Reasoning):** Reinforcement Learning to reward the use of `<think>` tags for logical tasks.
* **Stage 3: DPO (Persona):** Preference alignment to craft a friendly "Indian AI Assistant" identity.
* **Stage 4: Optimization:** Exported via **Unsloth** for high-efficiency inference.

---

## 📉 Known Limitations & Experimental Findings (The "Alignment Tax")
As an experimental 1.7B model, Indica demonstrates several critical findings regarding **Catastrophic Forgetting**:

* **Factual Regression:** Due to the limited parameter capacity, the final alignment (DPO) stage has caused the model to lose some precision in mathematical calculations and specific legal section numbering.
* **Persona Drift:** The model prioritizes its "creative persona" over technical accuracy. It may identify itself as an "AI Zindagi Manager" or other creative identities.
* **Logic Bypassing:** In some instances, the model may skip the internal `<think>` reasoning trace and provide direct, occasionally incorrect, answers.
* **Repetition Loops:** Occasional gibberish or repetition loops may occur in conversational Hinglish.

## 📦 Deployment for Testing
This model is best used to study **Hinglish conversational patterns** or as a base for further fine-tuning experiments.

### With Ollama
```bash
ollama run hf.co/prash616/Indica-1.7B-GGUF

```

## 🤝 Credits & Acknowledgements

- **Developer:** Prashant (`prash616`)
- **Base Model:** Alibaba Qwen Team
- **Training Framework:** Unsloth AI

### Disclaimer

This model is intended **solely for educational and research purposes**.  
It should **not** be used as a substitute for professional advice, including but not limited to **legal, agricultural, or mathematical decision-making**.