CortexDev — Neural Intelligence for Software Development

A focused, fine‑tuned code generation model designed to produce clean, executable Python code with minimal overhead.

---

✨ What is CortexDev?

CortexDev is a lightweight, fine‑tuned model based on Salesforce CodeGen‑350M using LoRA.
It’s built to think like a developer and output code‑only responses with consistent structure and style.

Key highlights

• ✅ Fine‑tuned model (not prompt engineering)
• ✅ Code‑only output by default
• ✅ Optimized for Colab & low‑VRAM environments
• ✅ Reproducible, professional fine‑tuning workflow

---

📸 Screenshots

Notebook + Gradio UI

Example generation

---

🧠 Why CortexDev?

Traditional prompt‑based assistants are inconsistent and verbose.
CortexDev embeds the behavior directly into model weights so you get:

Challenge	Traditional Prompting	CortexDev
Prompt dependency	Requires long system prompts	Built‑in behavior
Output consistency	Varies by prompt	Consistent code style
Extra explanations	Mixed output	Code‑only
Colab safety	Heavy models	Lightweight

---

🏗️ Architecture Overview

Base Code LLM (CodeGen‑350M)
        ↓
Frozen Base Weights
        ↓
LoRA Adapter Layers (Trainable)
        ↓
Custom Python Instruction Dataset
        ↓
CortexDev — Code Generation Model

---

🤖 Base Model

Salesforce CodeGen‑350M

• Pretrained on source code
• Strong syntax understanding
• Compact enough for Colab
• Ideal for LoRA fine‑tuning

Approx. VRAM usage: 6–8 GB (fp16 + LoRA)

---

🧩 Fine‑Tuning Strategy

Technique: LoRA (Parameter‑Efficient Fine‑Tuning)

✅ Frozen base weights
✅ Trainable adapter layers
✅ Lower memory usage
✅ Faster convergence

Method	Used
Training from scratch	❌
Full fine‑tuning	❌
LoRA / PEFT	✅

---

📦 Dataset Format

Each sample is structured as prompt → completion:

{
  "prompt": "Write a Python function to check if a number is prime.",
  "completion": "def is_prime(n):\n    if n <= 1:\n        return False\n    for i in range(2, int(n ** 0.5) + 1):\n        if n % i == 0:\n            return False\n    return True"
}

---

⚙️ Training Configuration (Colab‑Safe)

Parameter	Value
Batch Size	1
Epochs	2–3
Precision	fp16
Max Length	512
Optimizer	AdamW
Fine‑Tuning	LoRA

---

🧪 Inference Example

Input

Write a Python function to reverse a string

Output

def reverse_string(s):
    return s[::-1]

No explanations.
No markdown.
Just executable code.

---

🛠️ Tech Stack

• Python
• HuggingFace Transformers
• PEFT (LoRA)
• PyTorch
• Google Colab
• Gradio

---

📌 Use Cases

✅ Personalized coding assistants
✅ Hackathon accelerators
✅ Educational code generators
✅ Backend scaffolding tools
✅ Domain‑specific code models

---

✅ Final Note

CortexDev is not a chatbot.
It is a trained model that outputs code like a focused developer.