---
pretty_name: FireSafetyDataset Catalog
task_categories:
- object-detection
tags:
- fire
- smoke
- spark
- ember
- alarm
- hard-negative
---

# FireSafetyDataset catalog

This repository is a source catalog for a YOLO fire early-warning dataset. It contains no
training images. Source datasets remain independent until license review, label mapping,
deduplication and scene-leakage checks are complete.

## fire

Subcategories: `indoor`, `outdoor`, `forest`, `vehicle`.

- [firesafety-wildfire](https://huggingface.co/datasets/baizhanquan/firesafety-wildfire):
  `forest`; Apache-2.0; verified upstream labels are `0=smoke, 1=fire`.
- [firesafety-fire-smoke](https://huggingface.co/datasets/baizhanquan/firesafety-fire-smoke):
  mixed fire scenes; Apache-2.0; upstream labels are `0=fire, 1=smoke`.
- [fog_fire_detection bucket](https://huggingface.co/buckets/baizhanquan/fog_fire_detection-bucket):
  generic fire/smoke archive; `Fire` includes flames and embers, but there is no separate
  ember label. Its 40,900 images have passed basic local QA and detection normalization.
  Dataset metadata declares CC BY 4.0 while the README also mentions MIT, so release use
  remains blocked pending license clarification.

## smoke

Subcategories: `white_smoke`, `gray_smoke`, `black_smoke`.

- [firesafety-smoke-rf100](https://huggingface.co/datasets/baizhanquan/firesafety-smoke-rf100):
  CC BY 4.0; YOLO segmentation polygons; smoke color subtype is not verified.
- [firesafety-wildfire](https://huggingface.co/datasets/baizhanquan/firesafety-wildfire):
  wildfire smoke; color subtype is not verified.
- [firesafety-fire-smoke](https://huggingface.co/datasets/baizhanquan/firesafety-fire-smoke):
  mixed smoke scenes; color subtype is not verified.
- [fog_fire_detection bucket](https://huggingface.co/buckets/baizhanquan/fog_fire_detection-bucket):
  40,900-image generic fire/smoke source; color subtype is not verified and the license
  requires clarification.

## spark

Subcategories: `welding`, `grinding`, `electrical`.

No verified object-detection dataset has been accepted yet. Search terms should use
`welding sparks`, `grinding sparks` and `electrical arc sparks` rather than only `spark`.

## ember

Subcategories: `wildfire`, `campfire`.

No verified dataset has been accepted yet. Candidate searches should include `embers`,
`firebrands` and explicit flame-plus-ember label definitions. The Bucket's `Fire` label
includes embers, so it cannot train a separate ember class.

## alarm

Subcategories: `warning_light`, `smoke_detector`, `fire_alarm`.

No verified object-detection dataset has been accepted yet.

## negative

Subcategories: `cloud`, `fog`, `steam`, `sunset`, `lighting`, `exhaust`.

No verified hard-negative dataset has been accepted yet. The word `fog` in the Bucket name
does not describe a negative class; its `data.yaml` defines only `Fire` and `Smoke`.

Negative images are normally no-box training samples, not an additional YOLO class. The
current project-wide detector label standard remains `0=fire, 1=smoke`; any expansion to
spark, ember or alarm must be versioned as a separate label schema.

Machine-readable source revisions, mappings and coverage status are recorded in
[`catalog.yaml`](./catalog.yaml).