Revision	Training epochs	Best mAP50	Best mAP50-95
`hazard-objects-v1-full-v1`	100	0.8681	0.6555
`hazard-objects-v2-finetune-v1`	11	0.7721	0.5359
`hazard-objects-v2-finetune-v1-extend20`	20	0.8260	0.6409
`hazard-objects-v2-finetune-v2`	30	0.8322	0.6388
`hazard-objects-v2-finetune-v2-extend20`	8	0.8211	0.6202
`hazard-objects-v2-finetune-v3`	40	0.8381	0.6113
`hazard-objects-c270-finetune-v1`	20	0.8348	0.6382

The metrics above are taken from the epoch with the highest validation mAP50-95 in each run. Direct comparison requires evaluation on the same fixed test set.

Usage

python

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weight_path = hf_hub_download(
    repo_id="chocochip119/wardy-m05-hazard-detector",
    filename="weights/best.pt",
    revision="main",
)

model = YOLO(weight_path)
results = model.predict(
    source="example.jpg",
    imgsz=640,
    conf=0.25,
    save=True,
)

To load a specific release, replace main with a version tag such as hazard-objects-v2-finetune-v2.
Limitations
The model may produce false positives or miss hazardous objects.
Performance may decrease under poor lighting, occlusion, unusual camera angles, or unfamiliar environments.
The output should not be used as the sole basis for safety-critical decisions.
Additional validation and human review are required before real-world deployment.

License

A distribution license has not yet been specified. Verify the licenses and usage conditions of the model, base model, and training data before use or redistribution.