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

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

Indica-1.7B is a lightweight language model developed by Prashant to explore the limits of persona injection, cultural alignment, and reasoning behavior in ultra-small parameter architectures (1.7B).

Built on Qwen3-1.7B, the model was subjected to a rigorous post-training regime including Supervised Fine-Tuning (SFT), GRPO-based reasoning alignment, and Direct Preference Optimization (DPO).

---

🔬 Research Objective

This project investigates whether a 1.7B-parameter model can balance three traditionally competing objectives:

1. Domain Expertise
   Knowledge of Indian Law (IPC/BNS) and Agriculture.

2. Linguistic Persona
   Natural Hinglish/Hindi code-switching with colloquial Indian tone.

3. Logic & Reasoning
   Utilization of an explicit internal reasoning trace via native <think> tags.

---

🛠️ Post-Training Pipeline

The model underwent a specialized four-stage alignment strategy:

• Stage 1: SFT (Knowledge)
  Supervised fine-tuning on Indian Law and Agriculture datasets.

• Stage 2: GRPO (Reasoning)
  Reinforcement learning to reward structured reasoning using <think> tags.

• Stage 3: DPO (Persona Alignment)
  Preference optimization to shape a friendly, culturally grounded “Indian AI Assistant” identity.

• Stage 4: Optimization & Export
  Exported using Unsloth for efficient GGUF-based local inference.

---

📉 Known Limitations & Experimental Findings

(The “Alignment Tax”)

As an experimental 1.7B-parameter model, Indica exhibits several important alignment-related trade-offs:

• Factual Regression
  Due to limited parameter capacity, the final DPO stage introduces loss in precision for mathematical reasoning and exact legal section numbering.

• Persona Drift
  The model may prioritize its creative or conversational persona over strict technical accuracy, occasionally identifying itself as entities such as an “AI Zindagi Manager.”

• Logic Bypassing
  In some cases, the model may skip the internal <think> reasoning trace and respond directly, leading to incomplete or incorrect answers.

• Repetition Loops
  Occasional repetition or gibberish outputs may occur, particularly in long Hinglish conversations.

These behaviors are considered expected outcomes when aggressively aligning small models beyond their parameter limits.

---

📦 Deployment (For Testing & Research)

This model is best suited for:

• Studying Hinglish conversational behavior
• Exploring persona-alignment trade-offs
• Serving as a base for further fine-tuning experiments

Local Inference with Ollama

ollama run hf.co/prash616/Indica-1.7B-GGUF

🤝 Credits & Acknowledgements

• Developer: Prashant (prash616)
• Base Model: Alibaba Qwen Team
• Training Framework & Optimization: Unsloth AI

Disclaimer

This model is released strictly for educational and research purposes.
It should not be used for real-world legal, agricultural, or mathematical decision-making.

Indica-1.7B is an experimental exploration of how far cultural alignment and persona shaping can be pushed in small-scale language models—highlighting both their promise and their structural limits.