DevSh116/vastu-merged-llama3

VastuGPT – Fine-Tuned LLaMA-3 Vastu Shastra Expert

Model Description

VastuGPT is a fine-tuned large language model specialized in Vastu Shastra analysis, built by merging a domain-specific LoRA adapter into Meta-LLaMA-3-8B-Instruct.

The model is designed to provide clear, decisive, and actionable Vastu guidance and behaves as specified by system prompt

Unlike generic LLMs, VastuGPT avoids vague philosophy and focuses on practical remedies and rule-based decisions aligned with traditional Vastu principles.

It is trained on a highly diverse and accurate dataset of Various Vastu rules specifically for Apartments.

---

Model Details

• Base Model: Meta-LLaMA-3-8B-Instruct

• Fine-Tuning Method: LoRA (merged into base model)

• Architecture: Decoder-only Transformer

• Parameters: ~8 billion

• Quantization: INT8 compatible (bitsandbytes)

• Inference Engines Tested:

  • 🤗 Transformers
  • AirLLM (CPU/GPU offload)

---

Intended Use

Supported Use Cases

• Vastu consultation for:

  • Residential layouts
  • Room placement validation
  • Directional compliance checks

• Educational and research purposes

• Chatbot or API-based Vastu assistants

Not Intended For

• Legal, medical, or financial advice
• Structural engineering decisions
• Astrology, numerology, or horoscope predictions

---

Prompting Format

The model works best with a system-guided instruction format:

### System:
You are a strict and authoritative Vastu Shastra expert.
You clearly classify every placement as IDEAL, ACCEPTABLE, or INADVISABLE.
You always give practical remedies if something is wrong.

### User:
Is a toilet in the northeast acceptable?

### Response:

---

Example Outputs

Input: Is a toilet in the northeast acceptable?

Output: INADVISABLE — A toilet in the northeast disrupts the water and spiritual energy zone. Remedies: Relocate to northwest if possible; otherwise ensure constant ventilation, use sea-salt bowls weekly, and keep the area strictly clean.

---

Training Data

The model was fine-tuned using curated Vastu Shastra instructional data, including:

• Traditional directional rules
• Placement classifications
• Remedial guidelines

No personal, private, or sensitive user data was used.

---

Evaluation

The model was manually evaluated for:

• Consistency of classification (IDEAL / ACCEPTABLE / INADVISABLE)
• Deterministic decision-making
• Actionable remedies
• Reduced hallucination compared to base model

---

Limitations

• Outputs are based on learned textual patterns, not real-world verification
• Regional or school-specific Vastu interpretations may vary
• Should not replace professional architectural consultation

---

Ethical Considerations

• The model does not enforce beliefs
• Users are encouraged to treat outputs as guidance, not absolute truth
• No demographic or personal profiling is performed

---

License

This model inherits the license of the base model:

• License: Meta LLaMA 3 Community License
• Please review Meta’s terms before commercial use

---

Citation

If you use this model in research or applications, please cite:

@misc{vastugpt2026,
  title={VastuGPT: Fine-Tuned LLaMA-3 Model for Vastu Shastra},
  author={DevSh116},
  year={2026},
  url={https://huggingface.co/DevSh116/vastu-merged-llama3}
}

---

Author

Developed by DevSh116 🇮🇳 India