unicorn who dev commited on
Commit
cd244b2
·
verified ·
1 Parent(s): d8e1bf5

Update model catalogue and evidence boundaries

Browse files
Files changed (1) hide show
  1. README.md +18 -11
README.md CHANGED
@@ -13,28 +13,32 @@ FireViewer is an independent research and engineering project maintained by
13
 
14
  | Resource | Evidence-backed status |
15
  | --- | --- |
16
- | [D-FINE-M strict v1](https://huggingface.co/fireviewer/fire-smoke-dfine-m-strict-v1) | Primary research candidate. Independent COCO mAP50-95 0.3924; calibrated holdout F1 0.6475. Human review remains mandatory. |
17
- | [RT-DETRv2-R50 Fire/Smoke](https://huggingface.co/fireviewer/rtdetr-v2-r50-fire-smoke) | Conservative reference. Independent COCO mAP50-95 0.3502; lower false-alarm operating point. |
18
- | [YOLO11-M strict v1](https://huggingface.co/fireviewer/fire-smoke-yolo11m-strict-v1) | Published research baseline, not the primary detector. Independent COCO mAP50-95 0.1491; weak flame AP and recall are documented. |
19
- | [Strict-clean detection corpus v1](https://huggingface.co/datasets/fireviewer/fire-smoke-detection-corpus-v1) | 102,257 validated rows, source-specific provenance, 60,981/19,209/22,067 train/validation/test split. |
 
20
 
21
- The common independent benchmark uses 512 images from
22
  `Hajorda/flameye-wildfire-detection@test` at immutable revision
23
  `361a3dea8b877482af9f4ed80eff77ffd39926ef`. Its final selection has zero exact
24
- overlap with FireViewer train/validation hashes. Full protocol and limitations
25
- are included in the two new model cards.
26
 
27
  ## Other public resources
28
 
 
29
  - [Prithvi Burn Scars Training Dataset v1](https://huggingface.co/datasets/fireviewer/prithvi-burnscars-training-dataset-v1)
30
  - [DINOv3 Cross-View FireViewer v1](https://huggingface.co/datasets/fireviewer/dinov3-cross-view-fireviewer-v1-dataset)
31
  - [FireWarning Training Bundles v1](https://huggingface.co/datasets/fireviewer/firewarning-train-bundles-v1)
32
  - [Simple Measured Scenes v1](https://huggingface.co/datasets/fireviewer/simple-measured-scenes-v1)
33
  - [Dataset from Simulations](https://huggingface.co/datasets/fireviewer/dataset-from-simulations)
34
 
35
- Private calibration, holdout, unfinished DINOv3 and segmentation resources are
36
- not part of the public inventory. Deprecated, incomplete and superseded
37
- detector repositories were removed in the 5 September 2026 cleanup.
 
 
38
 
39
  ## Provenance, rights and integrity
40
 
@@ -45,7 +49,9 @@ before use or redistribution.
45
 
46
  Metrics are comparable only when dataset, split, immutable revision,
47
  preprocessing and evaluation protocol are identical. A model prediction is a
48
- probabilistic output, not geographic truth or publication authority.
 
 
49
 
50
  ## Code and contact
51
 
@@ -53,3 +59,4 @@ probabilistic output, not geographic truth or publication authority.
53
  - [Canonical documentation](https://github.com/fireviewer/Fireviewer_doc)
54
  - [Unicorn Who Dev](https://unicornwhodev.com)
55
  - Rights, collaboration and removal requests: unicornwhodev@gmail.com
 
 
13
 
14
  | Resource | Evidence-backed status |
15
  | --- | --- |
16
+ | [D-FINE-M strict v1](https://huggingface.co/fireviewer/fire-smoke-dfine-m-strict-v1) | Primary research candidate. Frozen independent COCO mAP50-95 0.3924; calibrated holdout F1 0.6475. Human review remains mandatory. |
17
+ | [RT-DETRv2-R50 Fire/Smoke](https://huggingface.co/fireviewer/rtdetr-v2-r50-fire-smoke) | Conservative historical reference. Frozen independent COCO mAP50-95 0.3502; lower false-alarm operating point. |
18
+ | [YOLO11-M strict v1](https://huggingface.co/fireviewer/fire-smoke-yolo11m-strict-v1) | Reproducible baseline, not the primary detector. Frozen independent COCO mAP50-95 0.1491; weak flame AP and recall are documented. |
19
+ | [RF-DETR Medium 1.10 strict v1](https://huggingface.co/fireviewer/fire-smoke-rfdetr-medium-v110-strict-v1) | Serious validation candidate (EMA mAP50-95 0.4683 at epoch 18), but **not yet evaluated on the frozen independent benchmark**. It must not be ranked against the rows above yet. |
20
+ | [Strict-clean detection corpus v1](https://huggingface.co/datasets/fireviewer/fire-smoke-detection-corpus-v1) | 102,257 automatically validated rows with source-specific provenance and a 60,981 / 19,209 / 22,067 train / validation / test split. |
21
 
22
+ The frozen independent detector benchmark uses 512 images from
23
  `Hajorda/flameye-wildfire-detection@test` at immutable revision
24
  `361a3dea8b877482af9f4ed80eff77ffd39926ef`. Its final selection has zero exact
25
+ overlap with FireViewer train/validation hashes. Only models whose cards report
26
+ that exact protocol may be compared using those independent scores.
27
 
28
  ## Other public resources
29
 
30
+ - [SegFormer B2 fire/smoke baseline v1](https://huggingface.co/fireviewer/segformer-b2-fire-smoke-baseline-v1)
31
  - [Prithvi Burn Scars Training Dataset v1](https://huggingface.co/datasets/fireviewer/prithvi-burnscars-training-dataset-v1)
32
  - [DINOv3 Cross-View FireViewer v1](https://huggingface.co/datasets/fireviewer/dinov3-cross-view-fireviewer-v1-dataset)
33
  - [FireWarning Training Bundles v1](https://huggingface.co/datasets/fireviewer/firewarning-train-bundles-v1)
34
  - [Simple Measured Scenes v1](https://huggingface.co/datasets/fireviewer/simple-measured-scenes-v1)
35
  - [Dataset from Simulations](https://huggingface.co/datasets/fireviewer/dataset-from-simulations)
36
 
37
+ Private calibration, holdout and retired training archives are not part of the
38
+ public catalogue. The DINOv3 pointing artifact is retained privately as a
39
+ five-epoch pilot because its positive localization evidence is too sparse for
40
+ public or operational claims. Deprecated, incomplete and superseded detector
41
+ repositories were removed in the 5 September 2026 cleanup.
42
 
43
  ## Provenance, rights and integrity
44
 
 
49
 
50
  Metrics are comparable only when dataset, split, immutable revision,
51
  preprocessing and evaluation protocol are identical. A model prediction is a
52
+ probabilistic output, not geographic truth or publication authority. Public
53
+ visibility does not imply deployment, qualification, rights clearance or
54
+ active use.
55
 
56
  ## Code and contact
57
 
 
59
  - [Canonical documentation](https://github.com/fireviewer/Fireviewer_doc)
60
  - [Unicorn Who Dev](https://unicornwhodev.com)
61
  - Rights, collaboration and removal requests: unicornwhodev@gmail.com
62
+