AOZGPT-LoRA-Adapters

AOZGPT: QLoRA Fine-tuned LLaMA-2 Assistant

AOZGPT is a helpful AI assistant created by fine-tuning LLaMA-2 7B using QLoRA (Quantized Low-Rank Adaptation) on the Dolly-15k instruction dataset. This model maintains the strong capabilities of LLaMA-2 while being optimized for helpful, truthful, and harmless interactions.

---

Model Details

Base Model

• Architecture: LLaMA-2 7B
• Model Type: Causal Language Model
• Fine-tuning Method: QLoRA (4-bit quantization + LoRA)
• Trainable Parameters: ~4 million (0.06% of total)
• Context Length: 512 tokens

Training Data

• Dataset: Databricks Dolly-15k
• Samples: 750 training examples + 150 validation examples
• Categories: Creative writing, Q&A, summarization, classification, information extraction, brainstorming

Fine-tuning Specifications

• Method: QLoRA with 4-bit NF4 quantization
• LoRA Rank: 8
• LoRA Alpha: 16.0
• Learning Rate: 2e-4
• Batch Size: 8 (effective 32 with gradient accumulation)
• Epochs: 1

---

Intended Use

AOZGPT is designed for:

• 🤖 AI Assistant Applications - Helpful, truthful responses to user queries
• 💬 Chatbots - Interactive conversations and Q&A
• 📚 Educational Tools - Explaining complex topics in simple terms
• 💻 Code Assistance - Writing and explaining code snippets
• 🎯 Instruction Following - Reliably executing user instructions

Direct Use

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load base model
model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load AOZGPT LoRA adapters
# (Implementation depends on your LoRA loading setup)

---

How to Use

Installation

pip install transformers torch accelerate bitsandbytes

Quick Start

from AOZGPT import ChatBot

# Initialize the chatbot
chatbot = ChatBot(
    model_name="meta-llama/Llama-2-7b-hf",
    adapter_path="./lora_adapters_final.pt"
)

# Single query
response = chatbot.chat("Explain quantum computing in simple terms.")
print(response)

# Interactive session
chatbot.interactive_chat()

Generation Parameters

response = chatbot.chat(
    "Your prompt here",
    max_new_tokens=256,    # Response length
    temperature=0.7,       Creativity (0.1-1.0)
    top_p=0.9,            # Diversity control
    top_k=50,             # Vocabulary restriction
    do_sample=True        # Enable sampling
)

---

Training Details

Hyperparameters

Training Results

• Final Training Loss: ~1.8
• Validation Loss: Consistent with training
• Perplexity: Improved throughout training
• Training Time: ~2 hours on A100 GPU

Prompt Format

The model expects conversations in this format:

system: You are AOZGPT, a helpful AI assistant...
user: [Your question or instruction]
assistant: [Model's response]

For context-aware responses:

system: You are AOZGPT...
context: [Additional context]
user: [Your question]
assistant: [Model's response]

---

Capabilities

✅ Instruction Following - Accurately follows user instructions
✅ Knowledgeable - Maintains LLaMA-2's broad knowledge base
✅ Helpful - Provides detailed, useful responses
✅ Truthful - Admits limitations and uncertainties
✅ Code Generation - Writes and explains code
✅ Creative Writing - Brainstorms and creates content
✅ Technical Explanations - Simplifies complex topics

Example Interactions

User: "Write a Python function to calculate Fibonacci numbers"
AOZGPT: "Here's a Python function to calculate Fibonacci numbers recursively..."

User: "Explain neural networks like I'm 10 years old"
AOZGPT: "Imagine a neural network like a team of tiny brain cells working together..."

User: "What are the benefits of renewable energy?"
AOZGPT: "Renewable energy has several key benefits including environmental sustainability..."

---

Limitations

• Context Window: Limited to 512 tokens
• Knowledge Cutoff: Inherits LLaMA-2's knowledge cutoff (2023)
• Occasional Hallucinations: May rarely generate inaccurate information
• Computational Requirements: Requires GPU for efficient inference
• Base Model Biases: Inherits any biases present in LLaMA-2

Bias, Risks, and Limitations

AOZGPT inherits the limitations and potential biases of the base LLaMA-2 model. Users should be aware that:

• The model may generate inaccurate or misleading information
• It should not be used for critical decisions without human verification
• Outputs may reflect biases present in the training data
• The model has no real-time knowledge beyond its training cutoff

Out-of-Scope Use

❌ Medical or legal advice
❌ Critical decision-making without verification
❌ Generating harmful or malicious content
❌ Real-time information beyond 2023

---

Technical Specifications

Model Architecture

• Base: LLaMA-2 7B Transformer
• Parameters: 7 billion (6.74B trainable with LoRA)
• Layers: 32
• Heads: 32
• Hidden Size: 4096
• Intermediate Size: 11008

Quantization

• Method: 4-bit NormalFloat (NF4)
• Double Quantization: Enabled
• Compute Dtype: bfloat16

LoRA Configuration

• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Rank: 8
• Alpha: 16.0
• Dropout: 0.05

---

Environmental Impact

• Hardware Type: A100 GPU (40GB)
• Training Time: ~2 hours
• CO2 Emissions: Estimated < 1 kg CO2eq
• Energy Consumption: ~2 kWh

Estimates based on ML CO2 Impact Calculator

---

License

This model is built on top of LLaMA-2, which is licensed for research and commercial use. Please refer to Meta's LLaMA-2 license for specific terms and conditions.

Note: The fine-tuned AOZGPT model inherits the license terms of the base LLaMA-2 model.

---

Contact

For questions, issues, or contributions:

• 💻 GitHub: https://github.com/AhmedZahran02

---

Acknowledgments

• Meta AI for the LLaMA-2 base model
• Databricks for the Dolly-15k dataset
• Hugging Face for the Transformers library
• The open-source community for QLoRA implementation

---

Disclaimer: This model is intended for research and educational purposes. Users are responsible for ensuring their use complies with applicable laws and regulations.