ExamNoteLLM: AI Universal Academic Assistant with QLoRA & RAG

An open-source academic intelligence platform combining QLoRA parameter-efficient fine-tuning (4-bit NF4) with a multi-tenant Retrieval-Augmented Generation (RAG) engine. Built to provide syllabus-grounded explanations, mark-specific exam notes, automated question generation, rubric-based student answer evaluation, and interactive revision across diverse subjects, institutions, and education levels.

---

Architecture Overview

                      +-----------------------------------+
                      |   Curriculum Knowledge Base       |
                      |   (PDFs, TXT, MD Syllabi/Notes)   |
                      +-----------------+-----------------+
                                        |
                                        v
+-----------------------+     +-------------------+     +-----------------------+
|  Student Query        | --> | RAG Vector Store  | --> | Prompt Construction   |
|  (Question, Marks,    |     | (MiniLM + Multi-  |     | (Syllabus Context +   |
|  Subject, University) |     | Tenant Metadata)  |     | Mark Rubric Guidance) |
+-----------------------+     +-------------------+     +-----------+-----------+
                                                                    |
                                                                    v
+-----------------------+     +-------------------+     +-----------+-----------+
| Multi-Turn Telegram   | <-- | Response Engine   | <-- | Llama-3-8B-Instruct   |
| Bot / Python CLI API  |     | (Streaming Token  |     | (4-bit QLoRA Adapter) |
|                       |     | Generation)       |     |                       |
+-----------------------+     +-------------------+     +-----------------------+

---

Core Capabilities

1. Syllabus-Grounded Explanations: Retrieves university syllabus modules and textbook references to keep model responses strictly within the course scope.
2. Mark-Constrained Answer Synthesis:
   • 2 Marks: Core definition and 2 bullet points (~180 tokens).
   • 5 Marks: Definition, 4-5 structured points, and a short example (~380 tokens).
   • 10 Marks: Full university structure: Definition, Mechanism, Key Properties, Real-World Application (~750 tokens).
   • 15 Marks: Comprehensive essay: Theory, Architecture, Comparative Table, Case Study, and Summary (~1100 tokens).
3. Automated Exam Question Generation: Generates balanced question papers (Part A definitions, Part B problem solving, Part C system design) complete with model answer keys.
4. Rubric-Based Student Answer Evaluation: Compares student submissions against authoritative textbook content to provide objective scorecards, key concepts covered, and missing points.
5. Study Materials & Adaptive Learning: Generates 5-minute pre-exam cheat-sheets, active-recall flashcards, and diagnostic 3-day recovery study plans.
6. Telegram Bot Interface: Multi-turn streaming bot with commands (/ask, /eval, /quiz, /notes, /profile).

---

Repository Structure

ExamNoteLLM/
├── notebooks/
│   └── ExamNoteLLM_Universal_Assistant.ipynb   # Complete end-to-end Colab notebook
├── src/
│   ├── config.py                               # Central hyperparameters and paths
│   ├── dataset.py                              # Dataset loaders and prompt formatting
│   ├── qlora_trainer.py                        # 4-bit quantization & SFTTrainer
│   ├── model_exporter.py                       # LoRA adapter merging utility
│   ├── rag_engine.py                           # SentenceTransformers vector index
│   ├── academic_assistant.py                   # Academic workflows (notes, eval, quiz)
│   └── telegram_bot.py                         # Multi-turn streaming Telegram assistant
├── data/
│   ├── sample_exam_questions.jsonl             # Benchmark evaluation dataset
│   └── academic_curricula/                     # Sample curriculum notes (DBMS, OS, ML)
├── scripts/
│   ├── train.py                                # CLI fine-tuning script
│   ├── evaluate.py                             # CLI evaluation script
│   └── run_bot.py                              # CLI bot launcher
├── requirements.txt                            # Dependencies
├── .gitignore
├── LICENSE
└── README.md

---

Quickstart Guide

1. Installation

git clone https://github.com/Adithyan435/ExamNoteLLM.git
cd ExamNoteLLM
pip install -r requirements.txt

2. Fine-Tuning with QLoRA

python scripts/train.py --dataset data/sample_exam_questions.jsonl --output_dir output/checkpoints

3. Running RAG Evaluation

python scripts/evaluate.py --question "Explain ACID properties in DBMS" --marks 10 --subject DBMS

4. Deploying Telegram Assistant

export TELEGRAM_API_KEY="your_bot_token_here"
python scripts/run_bot.py

5. Google Colab Training

Open notebooks/ExamNoteLLM_Universal_Assistant.ipynb directly in Google Colab with a free T4 GPU.

---

Training Hyperparameters

Hyperparameter	Value	Description
Base Model	Meta-Llama-3-8B-Instruct	8-billion parameter instruction-tuned model
Quantization	4-bit NormalFloat (NF4)	Memory reduction via BitsAndBytes
LoRA Rank (r)	16	Rank dimension for low-rank updates
LoRA Alpha	32	Scaling factor (2x rank)
Target Modules	q, k, v, o, gate, up, down	All attention and MLP projection layers
Optimizer	paged_adamw_8bit	Page memory optimizer for 16GB VRAM
Batch Size	2 per device (grad accum 8)	Effective batch size of 16
Learning Rate	2e-4 (Cosine decay)	Optimized for stable QLoRA convergence
Context Window	1024 tokens	Academic multi-mark answer capacity

---

License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.