som-tripathi-phi3-4k-indiangk

som-tripathi-phi3-4k-indiangk (Phi-3-Mini Indian History Expert)

som-tripathi-phi3-4k-indiangk is a lightweight, highly optimized, English-only fine-tuned model based on Microsoft's Phi-3-mini-4k-instruct architecture.

The model has been explicitly trained on a specialized, curation-grade dataset of structured historical facts covering ancient, medieval, and modern Indian history. Thanks to strict masking techniques applied during training, the model responds with direct, high-accuracy factual outputs while preventing repetition or conversation loops.

---

🚀 Key Features

• Localized Context Mastery: Deep knowledge of major Indian dynasties (Mauryas, Cholas, Mughals), key historical figures (Emperor Ashoka), the British Colonial era, landmark independence movements (Asahyog Andolan, Swaraj), and modern constitutional history (Presidents & Prime Ministers).
• Strict English-Only Anchor: Hardcoded system parameters ensure that the model consistently responds only in clean, native English.
• Ultra-Lightweight & Deployable: Provided in optimized 4-bit GGUF format (Q4_K_M), making it incredibly fast and perfect for running locally on weak hardware, laptops, or mobile devices via Ollama, LM Studio, or Llama.cpp.
• Overtraining Protection: Fine-tuned using unsloth with output-only response masking to prevent common factual hallucinations and question-repetition issues.

---

🛠️ Local Deployment with Ollama

You can easily pull this GGUF file and run it inside your local Ollama environment.

1. Create a Modelfile

Save the downloaded GGUF file (som-phi-gk.Q4_K_M.gguf) in a directory, and create a file named Modelfile (without any extension) next to it with the following configuration:

FROM ./som-tripathi-phi3-4k-indiangk.gguf

# Explicit ChatML Template configuration for Phi-3
TEMPLATE """{{ if .System }}<|system|>
{{ .System }}<|end|>
{{ end }}{{ if .User }}<|user|>
{{ .User }}<|end|>
{{ end }}<|assistant|>
{{ .Output }}<|end|>"""

# Hardstop parameters to prevent model spillover
PARAMETER stop "<|end|>"
PARAMETER stop "<|user|>"
PARAMETER stop "<|assistant|>"
PARAMETER stop "<|system|>"
PARAMETER stop "<|end_of_text|>"

# Conservative sampling for precise factual recovery
PARAMETER temperature 0.3
PARAMETER top_p 0.9

# Hardcoded English System Instructions
SYSTEM "You are a strict English-only assistant. You must answer accurately based on historical facts. Provide a direct, short answer to the user's question and stop immediately."

2. Build and Launch

Open your terminal in that folder and execute the following commands:

ollama create som-tripathi-phi3-4k-indiangk -f ./Modelfile
ollama run som-tripathi-phi3-4k-indiangk

---

📊 Training Specifications

• Base Foundation: unsloth/Phi-3-mini-4k-instruct (4-bit quantized)
• Hardware Used: Nvidia T4 GPU (via Google Colab)
• Optimization Framework: Unsloth (utilizing train_on_responses_only to mask user question loss calculations)
• Hyperparameters:
  • Rank (r): 16
  • Target Modules: All linear layers (q, k, v, o, gate, up, down_proj)
  • Learning Rate: 2e-4
  • Batch Size: 4 (with 2 Gradient Accumulation Steps)
  • Total Steps: 30 (optimized to avoid catastrophic forgetting)

---

📜 License & Acknowledgements

This project is open-sourced under the MIT License. Special thanks to the Microsoft Phi team for the incredibly efficient base model architecture and the Unsloth AI team for providing the core libraries to execute memory-safe local micro-fine-tuning.