This model is a specialized fine-tuned version of Qwen/Qwen2-VL-7B-Instruct designed for the automated inspection of telecommunication infrastructure.

Developed as a graduation project for the Digital Egypt Pioneers Initiative (DEPI) - GenAI Track.

Model Description

The model is trained to identify and evaluate the health of telecom assets such as Antennas, RRUs (Remote Radio Units), and Battery Cabinets. It uses visual information to detect issues like rust, physical damage, and equipment degradation.

Architecture: Qwen2-VL-7B with LoRA adapters.
Fine-tuning Technique: PEFT (LoRA).
Task: Multimodal Image-to-Text (Vision-Language).

Intended Uses & Limitations

Intended Uses:

Automated site surveys for telecom engineers.
Preliminary health checks for equipment in remote locations.
Educational tool for identifying telecom hardware.

Limitations:

The model may struggle with highly occluded images or extreme low-light conditions.
Accuracy depends on the quality of the input image; 384x384 resolution is recommended for optimal balance.

Training Data

The model was trained on a curated dataset of telecom equipment images, including both healthy and damaged states. The dataset focuses on:

Antennas: Identifying rust and structural alignment.
RRUs: Detecting cable damage and physical impact.
Batteries: Checking for leakage or cabinet integrity.

Training Procedure

Training Hyperparameters

The following hyperparameters were used during training:

Learning Rate: 1e-4
Optimizer: AdamW
Batch Size: 1 (Per device)
Gradient Accumulation Steps: 4
Epochs: 3
Mixed Precision: FP16
Quantization: 4-bit (bitsandbytes)

Training Results

Final Training Loss: ~10.49.
Hardware: 2x NVIDIA T4 GPUs on Kaggle.

Framework Versions

PEFT 0.18.1
Transformers 5.3.0
Pytorch 2.9.0+cu126

David0dods/Qwen-telecom-asset-model

Qwen-telecom-asset-model