rabahdev/fire-smoke-yolov8n

YOLOv8n — Fire & Smoke Detector (D-Fire fine-tune)

YOLOv8n fine-tuned on the D-Fire dataset for detecting smoke and fire in images.

Classes

• 0 — smoke
• 1 — fire

Files

• best.pt — fine-tuned weights (use this)
• last.pt — final-epoch weights

Results (test split, 4,306 images)

Metric	All	Smoke	Fire
mAP50	0.754	—	—
mAP50-95	0.430	0.499	0.362
Precision	0.766	—	—
Recall	0.688	—	—

Training config

• Base: yolov8n.pt (COCO pretrained)
• Epochs: 50, image size: 640, batch: 16
• Optimizer: MuSGD (auto), lr0=0.01
• Device: Apple MPS

Usage

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

ckpt = hf_hub_download(repo_id="rabahdev/fire-smoke-yolov8n", filename="best.pt")
model = YOLO(ckpt)
results = model("image.jpg")
results[0].show()

Dataset

• D-Fire (smoke + fire detection), YOLO format
• Train: 14,122 — Val: 3,099 — Test: 4,306

License

AGPL-3.0 (inherited from Ultralytics YOLOv8).