Instructions to use fireviewer/yolo11m-fire-smoke with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- ultralytics
How to use fireviewer/yolo11m-fire-smoke with ultralytics:
from huggingface_hub import hf_hub_download from ultralytics import YOLO # pick the weights file from this repo's "Files and versions" tab weights = hf_hub_download("fireviewer/yolo11m-fire-smoke", "<weights>.pt") model = YOLO(weights) source = 'http://images.cocodataset.org/val2017/000000039769.jpg' model.predict(source=source, save=True) - Notebooks
- Google Colab
- Kaggle
YOLO11M · Historical source checkpoint
Historical conversion-source checkpoint. Class order: 0=fire, 1=smoke. RGB input with Ultralytics letterboxing at 928 px.
Checkpoint and status
best.pt is the retained published checkpoint. This distinct 42-epoch checkpoint is retained as the historical source of the published Cadryl conversion and for reproducibility. It is not the recommended current YOLO11M detector. Its weights differ from the improved A100 checkpoint; it is not an exact duplicate.
Validation evidence
Best recorded AP50–95: 38.6300%, epoch 42. The table below describes that E42 measurement.
| Class | Precision | Recall | F1 | AP50 | AP75 | AP50–95 | Reference boxes |
|---|---|---|---|---|---|---|---|
| Global | 67.51% | 61.98% | 64.63% | 65.07% | 38.89% | 38.63% | 10,925 |
| fire | 65.26% | 59.98% | 62.51% | 62.56% | 37.53% | 36.83% | 5,727 |
| smoke | 69.77% | 63.98% | 66.75% | 67.58% | 40.26% | 40.44% | 5,198 |
Validation contains 11,266 images and 10,925 reference boxes. AP50–95 averages IoU thresholds 0.50 to 0.95. Operating-point precision and recall are different measurements from AP. No new benchmark or held-out test score is produced by this presentation.
Annotation example and inference
The image shows source dataset annotations. It is not an inference result or a model-accuracy demonstration. Coordinate and drawing guide.
pip install ultralytics huggingface_hub
from huggingface_hub import hf_hub_download
from ultralytics import YOLO
weights = hf_hub_download(
repo_id="fireviewer/yolo11m-fire-smoke",
filename="best.pt",
revision="ecf4ca351b5dbd4ab84a92c0402e6f45f8001de9",
)
model = YOLO(weights)
results = model.predict("image.jpg", imgsz=928, conf=0.25)
for index, result in enumerate(results):
result.save(filename=f"prediction-{index:03d}.jpg")
for box in result.boxes:
class_id = int(box.cls.item())
print(model.names[class_id], box.xyxy[0].tolist(), box.conf.item())
The code saves predicted boxes to prediction-000.jpg. Output boxes are [x1, y1, x2, y2] in source-image pixels. conf=0.25 is an example threshold to tune for the target domain. It differs from the reference-annotation image shown above. The original run used Ultralytics 8.4.210.
Data and run parameters
Dataset: fireviewer/fireviewer_fire_smoke_flame_v3_native, revision 8c8b197fa1c928217529a04bff434fb70a66614e. Training / validation / test: 78,450 / 11,266 / 13,523 images. This is the training snapshot, not the current dataset population. Retain its native annotations and split membership. Experimental DINO/SAM/Qwen or crop-audit proposals are not applied by this card.
| Recorded parameter | Value |
|---|---|
| Initialization | COCO-pretrained yolo11m.pt |
| GPU | NVIDIA A100-SXM4-80GB |
| Resolution | 928 px, Ultralytics letterboxing |
| Physical batch / nominal batch | 55 / nbs=64 |
| Optimizer / initial LR | AdamW / 0.0005 |
| Schedule | Cosine decay; 5 warmup epochs |
| Weight decay / precision | 0.0005 / AMP |
| Augmentations | Mosaic, MixUp and Copy-Paste disabled; scale 0.3 |
| Stop | Manual after 42 complete epochs; configured patience 7 |
Limitations and licence
Faint or distant smoke, small fires, reflections, steam, occlusion and camera-domain changes can produce omissions and false positives. These models do not establish the cause, geographic origin or actual danger of a fire. Validation scores do not qualify an autonomous emergency alarm.
Declared model licence: AGPL-3.0. Source-image conditions are separate and remain attached to the dataset. Presentation changes do not strip or convert the retained checkpoint; its original embedded contents remain intact. Keep the model, labels, preprocessing and dataset revisions together.
FireViewer · Resource catalogue · Presentation updated 14 November 2026.
- Downloads last month
- 350
Dataset used to train fireviewer/yolo11m-fire-smoke
Collection including fireviewer/yolo11m-fire-smoke
Evaluation results
- mAP@0.50:0.95 (%) on FireViewer · validationvalidation set self-reported38.630
- mAP@0.50 (%) on FireViewer · validationvalidation set self-reported65.070

