--- configs: - config_name: default data_files: - split: train path: data/train-* - split: validation path: data/validation-* - split: test path: data/test-* dataset_info: features: - name: image dtype: image: decode: false - name: sha256 dtype: string - name: width dtype: int32 - name: height dtype: int32 - name: split_group dtype: string - name: merge_role dtype: string - name: objects struct: - name: bbox list: list: float32 length: 4 - name: category list: int32 splits: - name: train num_bytes: 26810502819 num_examples: 78455 - name: validation num_bytes: 2096247471 num_examples: 11266 - name: test num_bytes: 2343043141 num_examples: 13523 download_size: 31254424598 dataset_size: 31249793431 pretty_name: FireViewer · Probable fire v3 · native images task_categories: - object-detection tags: [fireviewer, fire, smoke, probable-fire, native-images] --- ![FireViewer · Probable fire v3](assets/card-header.svg) # Probable fire v3 A **103,244-image native corpus** for proposing visible fire-or-smoke regions under a single `probable_fire` class. It is separate from [Fire & Smoke v3](https://huggingface.co/datasets/fireviewer/fireviewer_fire_smoke_flame_v3_native) and does not automatically receive that corpus’s later additions. ![Published split counts and the shared probable_fire class](assets/dataset-overview.png) | Split | Images | |---|---:| | `train` | 78,455 | | `validation` | 11,266 | | `test` | 13,523 | ## Choose the right task | Need | Dataset | Classes | |---|---|---| | Propose regions containing visible fire or smoke | This repository | `0=probable_fire` | | Distinguish and locate flame and smoke | Fire & Smoke v3 | `0=flame`, `1=smoke` | The two corpora have different populations and revisions. Do not transfer their category IDs or annotations without an explicit conversion. ## Schema and geometry Native images are embedded in `data/*.parquet`. `sha256` identifies source bytes; `width` and `height` give native dimensions. `split_group` retains grouping and `merge_role` records the merge role. `objects.bbox` stores `[x, y, width, height]` in native pixels. `objects.category` uses **0=probable_fire**; the two lists align by index. The common class does not distinguish flame from smoke or certify an actual incident. It does not use the older v2.5 COCO category IDs. ![Schematic XYWH-to-XYXY conversion](https://huggingface.co/datasets/fireviewer/fireviewer_fire_smoke_flame_v3_native/resolve/main/assets/box-coordinates.png) ## Streaming example Install `datasets` and `pillow`, then load a pinned revision: ```python import io from datasets import load_dataset, Image as HFImage from PIL import Image ds = load_dataset( "fireviewer/fireviewer_probable_fire_v3_native", revision="637ddffba5442991ec4110283f1b3e49478f1a62", split="validation", streaming=True, ).cast_column("image", HFImage(decode=False)) sample = next(iter(ds)) image = Image.open(io.BytesIO(sample["image"]["bytes"])) print(image.size, sample["objects"]) ``` ## Evidence and rights Counts come from the published metadata. The documentation review inspected the Parquet schema and an annotation sample; it is not exhaustive human review or proof of event-level independence. This repository does not contain a blanket image licence or a detailed standalone provenance dossier. Public access does not establish redistribution rights. Check original sources and their conditions for the intended use. Detections remain visual estimates. [FireViewer](https://huggingface.co/fireviewer) · [Resource catalogue](https://huggingface.co/spaces/fireviewer/README/blob/main/docs/RESOURCE_INDEX.md) · Presentation updated **14 November 2026**.