llama3-ayurveda-text-v4

🧠 Model Name: llama3-ayurveda-text-v4
📚 Base Model: unsloth/Llama-3.2-11B-Vision-Instruct-bnb-4bit
🧪 Trained By: AyurEze 🏥 Domain: Ayurveda / Healthcare / Medical
📍 Language(s): English, Tamil (bilingual capability for Ayurveda-related queries)
🧵 Quantization: 4-bit (bnb)
🧩 Framework: Transformers, PEFT, Unsloth

---

🧾 Model Description

This is a fine-tuned LoRA version of LLaMA 3.2 11B (bnb 4-bit) adapted for Ayurvedic medicine and natural language generation in the healthcare domain, with special emphasis on traditional Indian medical knowledge.

It is trained on a curated dataset of 110+ classical Ayurvedic drugs and herbs, slokas from Bhavaprakasha, Bhaisajya Ratnavali, and traditional usage explanations in both English and Tamil. The goal is to provide accurate, contextually grounded, and culturally sensitive answers to questions about Ayurveda.

---

💡 Use Cases

• 🧪 Integrate with RAG pipelines for Ayurveda apps like AyurEze
• 🤖 Q&A assistant for traditional medicine consultations

---

🛠️ Training Details

• Base Model: unsloth/Llama-3.2-11B-Vision-Instruct-bnb-4bit (text-only use)
• Adapter Type: LoRA (using PEFT)
• Library: unsloth, transformers, peft
• Quantization: 4-bit bnb
• Epochs: 1–3 depending on dataset size
• Batch Size: 2 per device
• Data Format: Instruction-tuned, chat-style prompts with messages key ({"messages": [...]})

---

📊 Evaluation

Metric	Result
Human Eval	✅ Relevant, accurate for domain-specific queries
Token Quality	✅ Well-structured output for Ayurvedic prompts
Toxicity	✅ Low (clean dataset, domain-specific)

---

🔍 Sample Prompt

{
  "messages": [
    { "role": "user", "content": "Explain the uses of Pippali in Ayurveda in simple Tamil." }
  ]
}