abhijitbetigeri/dc-ops-dataset
DC-Ops: Data Center Components Dataset On-device data center operations assistant dataset for the Qualcomm x Meta ExecuTorch Hackathon. Overview 319 images of data center infrastructure (server racks, NVIDIA NVL72, cables, ports, etc.) 3,045 polygon annotations in YOLO-seg format 16 component classes auto-labeled with Grounding DINO + SAM, for fine-tuning YOLOv8n-seg Classes ID Class Count 0 server rack rack enclosures 1 compute tray… See the full description on the dataset page: https://huggingface.co/datasets/abhijitbetigeri/dc-ops-dataset.
DC-Ops: Data Center Components Dataset
On-device data center operations assistant dataset for the Qualcomm x Meta ExecuTorch Hackathon.
Overview
- 319 images of data center infrastructure (server racks, NVIDIA NVL72, cables, ports, etc.)
- 3,045 polygon annotations in YOLO-seg format
- 16 component classes auto-labeled with Grounding DINO + SAM, for fine-tuning YOLOv8n-seg
Classes
Trained Model
models/dc_ops_yolov8n_seg.pt — YOLOv8n-seg fine-tuned on this dataset (6.5MB)
- Box mAP50: 0.232
- Top classes: compute tray (0.57), network port (0.51), cable (0.27)
- Target device: Samsung Galaxy S25 Ultra (Snapdragon 8 Elite, SM8750)
Label Format
YOLO-seg format: class_id x1 y1 x2 y2 ... xn yn (normalized polygon coordinates)
Pipeline
Brightdata scraping → Grounding DINO + SAM auto-labeling → YOLOv8n-seg fine-tuning → ExecuTorch .pte export → Snapdragon NPUUsage
from ultralytics import YOLO
model = YOLO("models/dc_ops_yolov8n_seg.pt")
results = model("path/to/server_rack_image.jpg")