006aman/Bharat-Model-V1

Bharat: Indian Knowledge Systems AI (V1)

Bharat is a custom-trained, 7-billion parameter language model fine-tuned specifically on Indian Knowledge Systems (IKS). It is designed to go beyond standard Wikipedia summaries and provide deep, culturally authentic insights into Indian history, Ayurveda, classical arts, and ancient heritage.

🛠️ Model Details

• Base Model: Mistral 7B (Mistral-7B-v0.1 base)
• Training Framework: Unsloth & Hugging Face
• Dataset: 15,001 curated instruction pairs distilled from 286 IKS texts using Gemini API.
• Format: 4-bit Quantized GGUF (q4_k_m) for highly efficient local CPU/GPU inference.

⚠️ Known Limitations (V1)

Extensive real-world testing of Bharat V1 revealed several behavioral quirks due to dataset and training configuration imbalances:

• Chat Template Mismatch (Critical): During training, the instruction data was formatted using Llama 3 special tokens (<|begin_of_text|>, <|eot_id|>) on a Mistral 7B base. Because Mistral's tokenizer does not natively map these as control tokens, V1 does not recognize normal stopping boundaries.
• Format Amnesia (Self-dialogue): Multi-turn conversations were packed together, causing the model to generate the user's subsequent questions instead of stopping (e.g., typing "hii" causes it to write out a full question like "What is the Kumbh Mela?" and answer it itself).
• Over-Storytelling: The model heavily rewards poetic, sensory imagery. It struggles to provide concise, direct answers to simple utility questions (e.g., answering a math or programming question with a flowery philosophical story).
• Instruction Obedience: V1 struggles with strict formatting constraints (like JSON output or one-word replies) due to a lack of contrastive examples in the corpus.
• Runaway Responses: V1 frequently ends answers by inviting further discussion, leading to unnatural conversational loops.

---

🚧 Roadmap for Bharat V2 (In Progress - Data Phase Completed)

A completely redesigned training pipeline is currently under development. The V2 dataset preparation, quality cleaning, and audits are 100% complete, and the dataset is ready for the upcoming fine-tuning run.

The V2 improvements include:

1. Chat Template Alignment: Training will adopt Mistral's native format (<s>[INST]...[/INST]</s>) or align the base model to prevent tokenizer and stop-token errors.
2. System Prompt Integration: The core "Bharat" guide persona and behavioral constraints are baked directly into every single training example.
3. Conversational Balance: Includes a new Greetings & Utility partition (10%) to handle simple introductions and non-IKS tasks, and a Knowledge Calibration partition (2%) to train refusal/hedging behaviors for uncertain or debated historical claims.
4. Data Hygiene: 758 first-person memory hallucinations rewrote to third-person objective accounts; 109 fabricated academic citations stripped; gravity concepts refined to ākarṣaṇa-śakti; and Aryabhata heliocentrism claims corrected to axial rotation.

---

🚀 How to Run V1 Offline (Mac/Windows/Linux)

Because of the Llama 3 chat template mismatch on the Mistral base model, you MUST configure Ollama to use Llama 3 special tokens as stop parameters. Otherwise, the model will experience runaway generation and self-dialogue loops.

1. Download the Weights

Download the iks-mistral-7b-q4_k_m.gguf file from the Files and versions tab in this repository.

2. Create the Modelfile

Create a file named Modelfile in the same directory as the GGUF model and add the following configuration:

cat << 'EOF' > Modelfile
FROM ./iks-mistral-7b-q4_k_m.gguf

SYSTEM "You are Bharat, an AI assistant specialized in Indian Knowledge Systems (IKS). You have deep knowledge of Ayurveda, Yoga, Indian philosophy, ancient architecture, classical music, mathematics, astronomy, and cultural heritage. Answer questions thoughtfully and accurately. If asked something outside IKS, gently redirect to a relevant Indian knowledge topic."

TEMPLATE """<|begin_of_text|>{{ if .System }}<|start_header_id|>system<|end_header_id|>

{{ .System }}<|eot_id|>{{ end }}<|start_header_id|>user<|end_header_id|>

{{ .Prompt }}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

"""

PARAMETER stop "<|eot_id|>"
PARAMETER stop "<|start_header_id|>"
PARAMETER temperature 0.7
PARAMETER top_p 0.9
EOF

3. Build and Run in Ollama

Run these commands in your terminal to build and launch the model:

# Create the local model in Ollama
ollama create Bharat -f Modelfile

# Run the model
ollama run Bharat