ask-m-gemma3-exam-lora

Ask-M Gemma-3 Exam LoRA

This repository contains a LoRA adapter fine-tuned on exam-style question answering data for the Gemma-3 12B model.

The adapter is designed to generate concise, marks-aware answers suitable for university examinations.

---

🎯 Purpose

This LoRA specializes the base model to:

• Write answers appropriate to given marks
• Be clear, correct, and concise
• Follow exam-oriented structure
• Avoid unnecessary elaboration or follow-up questions

It is intended for use in automated academic assistance systems (e.g., Ask-M).

---

🧠 Training Overview

• Base model: unsloth/gemma-3-12b-it-bnb-4bit
• Fine-tuning method: QLoRA
• Trainable parameters: ~24M (~0.2% of base model)
• Sequence length: 1024
• Epochs: 1
• Optimizer: paged_adamw_8bit
• Hardware: NVIDIA T4 (Kaggle)

---

📂 Dataset Format

Each training example follows this structure:

{
  "instruction": "You are an exam-answering assistant...",
  "input": "Subject: PHYS101\nMarks: 5\nQuestion: ...",
  "output": "Structured exam-style answer...",
  "meta": {
    "subject": "PHYS101",
    "marks": 5,
    "keywords": ["..."]
  }
}