Khalidahmad01/cortex-copilot-lora

Cortex Copilot LoRA Adapter

Overview

This repository contains the LoRA adapter developed for the Cortex Copilot Engineering Challenge.

The adapter fine-tunes Qwen2.5-3B-Instruct using QLoRA to improve responses for industrial energy management, electrical engineering concepts, Indian electricity tariff logic, and Cortex-specific metrics.

---

Base Model

Qwen/Qwen2.5-3B-Instruct

---

Fine-Tuning Method

• Framework: Unsloth
• Method: QLoRA
• PEFT (Parameter-Efficient Fine-Tuning)

---

Dataset

The model was fine-tuned on 399 instruction-response pairs covering:

• Electrical Engineering Concepts
• Indian Industrial Tariff Rules
• Cortex Metric Explanations
• Energy Optimization
• Refusal Behaviour
• Tenant Isolation
• Prompt Injection Resistance

---

Training Configuration

Parameter	Value
Epochs	3
Batch Size	2
Gradient Accumulation	4
Learning Rate	2e-4
Max Sequence Length	2048
GPU	Tesla T4

---

Files

• adapter_model.safetensors
• adapter_config.json
• tokenizer.json
• tokenizer_config.json
• chat_template.jinja

---

Intended Use

This adapter is intended for educational purposes as part of the Cortex Copilot Engineering Challenge.

It specializes the base model for industrial energy management while relying on external telemetry data for real-time information.

---

Author

Khalid Ahmad Raza

B.Tech Computer Engineering

National Institute of Technology Kurukshetra