bhriguverma/padhai-v2-final-gguf

PadhAI v2 Final: Multimodal Punjabi AI Tutor (Gemma 4)

PadhAI v2 Final is a specialized, fine-tuned multimodal AI tutor based on the Gemma 4 4B (E4B) architecture. It is purpose-built to support students of the Punjab School Education Board (PSEB) by providing explanations, answering questions, and clarifying concepts from the official curriculum in fluent Punjabi (Gurmukhi).

Developed by Bhrigu Verma.

🌟 Key Features

• Gemma 4 Architecture: Leverages the latest efficient 4-billion parameter model for high-speed reasoning and deep multilingual understanding.
• PSEB Curriculum Grounded: Trained specifically on cleaned datasets including Agriculture, Science, and Mathematics textbooks from the Punjab School Education Board.
• Punjabi First: Optimized for natural, warm, and culturally relevant responses in Punjabi (Gurmukhi script).
• Multimodal Capabilities: Can be paired with a vision projector to analyze textbook images, diagrams, and handwritten notes.
• Mobile Ready: Distributed in GGUF format for local execution on mobile devices (Android/iOS) and resource-constrained hardware.

📊 Training Details

• Base Model: google/gemma-4-E4B-it
• Dataset: Exclusively trained on aafinal/train_data.jsonl (2,851 high-quality instructional pairs).
• Hardware: Trained on NVIDIA H100 (HBM3) using Unsloth for 2x faster optimization.
• Optimization: BF16 precision with QLoRA (Rank 64, Alpha 128).
• Final Metrics:
  • Train Loss: 2.045
  • Eval Loss: 1.473
  • Epochs: 2

📁 Files in this Repository

• padhai-v2-final-q4_k_m.gguf: The language model weights quantized to 4-bit (Q4_K_M).
• adapter_model.safetensors: The raw LoRA adapter weights (for researchers/developers).
• train_data.jsonl: The exact dataset used for the final retraining session.
• training_log.txt: Comprehensive logs detailing the loss curves and throughput.
• training_config.json: The hyperparameters used for this specific run.

🚀 How to Use

Mobile (PocketPal / LM Studio Mobile)

1. Download padhai-v2-final-q4_k_m.gguf.
2. Load it in your preferred GGUF-compatible app.
3. (Optional) For vision support, use the padhai-mmproj-f16.gguf vision tower.

CLI (llama.cpp)

./llama-cli   -m padhai-v2-final-q4_k_m.gguf   --mmproj padhai-mmproj-f16.gguf   -p "ਤੁਸੀਂ ਕੌਣ ਹੋ?"   -n 512

⚠️ Limitations & Scope

• Curriculum Scope: PadhAI is designed for PSEB school topics. It may refuse or provide generic answers for advanced university-level or non-educational topics.
• Hallucinations: While improved, the model may occasionally generate incorrect information. Always cross-reference with official textbooks.

📜 Credits & License

• Lead Developer: Bhrigu Verma
• Base Model: Google Gemma 4
• Optimization: Unsloth AI
• License: Released under the Gemma Terms of Use.

---

Created with ❤️ for the students of Punjab.