Datasets:
License:
| pretty_name: FireViewer Training Bundles v1 | |
| license: other | |
| language: | |
| - en | |
| - fr | |
| size_categories: | |
| - 100G<n<1T | |
| task_categories: | |
| - object-detection | |
| - image-segmentation | |
| - text-generation | |
| - visual-question-answering | |
| tags: | |
| - wildfire | |
| - fire-detection | |
| - burned-area | |
| - cross-view-localization | |
| - geospatial | |
| # FireViewer Training Bundles v1 | |
| Les identifiants FireWarning présents dans le slug, les noms d’archives, les manifestes, les stages et les empreintes sont conservés pour la compatibilité et la traçabilité. Ils ne désignent pas un projet actif distinct. | |
| Fifteen self-contained, versioned ZIP64 packages prepared for the FireViewer training pipeline. | |
| Every training objective has exactly one ZIP. Each ZIP contains its source payloads, | |
| manifests, attribution and license metadata, entry-point contract, and a complete SHA-256 | |
| inventory under a single root directory. | |
| > **Contrats à ne pas exécuter tels quels.** Les ZIP publics | |
| > `fire-pointing-lora-v1.zip` et `cross-view-localization-v1.zip` conservent leur ancien | |
| > `TRAIN_BUNDLE.json`. Les données et | |
| > empreintes restent valides, mais les entrypoints doivent être remplacés par la révision | |
| > `mvp-a40-v2` décrite dans les manifestes versionnés du dépôt. Aucune reconstruction ni | |
| > republication de ces deux ZIP n'est effectuée dans cette passe. | |
| | Training objective | ZIP | Size (bytes) | SHA-256 | State | | |
| |---|---|---:|---|---| | |
| | Fire/smoke media filter | `media-filter-fire-smoke-v1.zip` | 32,772,718,037 | `ed6ee4cf71a2538cfce7b95e6e548528d94b4cea095c8d03ccfc4c0e5ba4c8ae` | Training-ready | | |
| | Burned-area segmentation | `burned-area-segmentation-v1.zip` | 27,926,097,197 | `85c2f17248528ebbd5aa8395e72435ba5a12626bb5a53f5730109b11ea5dde36` | Dataset-ready; training blocked until independent geographic test | | |
| | Cross-view localization | `cross-view-localization-v1.zip` | 15,895,751,194 | `5f5e14083da209bd978117e2c7470f8f63decdb4ba3223914da7e7164aa64f5f` | Data valid; DINOv2 entrypoint obsolete, rebuild required | | |
| | Fire pointing LoRA | `fire-pointing-lora-v1.zip` | 32,589,764,131 | `a43cda497078b89960c300fce26280171441cf6111c208f96adfcfdecdc9762b` | Data valid; Qwen entrypoints obsolete, rebuild required | | |
| | Cross-view registration | `cross-view-registration-v1.zip` | 2,299,599,606 | `a7e7d9205c34f96cf1843a7b7a9455e81a704c2926b9ba648bb00ecbddbc2f1e` | Dataset-ready; RoMa quality gate, double-validated test, and fine-tuning contract pending | | |
| | Fire/smoke/normal media triage | `media-triage-fire-smoke-v1.zip` | 31,688,094,244 | `0ce5a30979007468efa8dd2f2ec5b20547943e080a6ed36318297023456ece34` | Dataset-ready; classifier trainer and independent critical test pending | | |
| | Orchestrator gates SFT | `orchestrator-gates-sft-v1.zip` | 17,444 | `3ca4813cae6f04b806fce16be7a08a6854409dcd96df166d62a346e6d9a809ff` | Dataset-ready; SFT trainer and independent human validation pending | | |
| | Fire progression/front inference | `fire-progression-front-inference-v1.zip` | 6,493,320 | `0bb682ee726d3ec62054e5c44b7dc84e0a0658e5e8ca5a1286662056020b7f11` | Dataset-ready; observed active-front labels and independent geographic test pending | | |
| | Daily wildfire fact synthesis | `daily-wildfire-fact-synthesis-v1.zip` | 222,162 | `9e4c7dde666b964a3f38ae64c50383cd7faff7fbdab9d32a42b6ccca398fe332` | Dataset-ready; raw source documents and France validation pending | | |
| | Structured wildfire situation/resources | `wildfire-situation-resources-structured-v1.zip` | 8,696,884 | `fc8bd5f419183b1ba603aab2117be842b75ad9acd46f9ba10eb2fff73e6ceb02` | Dataset-ready; France validation and stable incident identifiers pending | | |
| | Engaged-assets object detection | `engaged-assets-object-detection-v1.zip` | 1,193,478,095 | `9a8e746f77ed88b7a8b3fa29ea13448f12ee9e8cdc6841ddfd8175950c079178` | Dataset-ready; fire-engine and operational-role labels pending | | |
| | Camera depth and pose prior | `camera-depth-pose-prior-v1.zip` | 39,593,395,922 | `4f9574b2920126e7f34d62759ff0192c53062bc337e206309106beb0f670c280` | Dataset-ready; synthetic-only, real rural critical test pending | | |
| | Outdoor metric depth | `outdoor-metric-depth-v1.zip` | 132,759,914,158 | `b9aa757c2e52007a3db40fc3c7f0878fb7dd391464832142e8d9b104cbb22bf2` | Training-ready as an auxiliary real outdoor depth prior | | |
| | Wildfire smoke detection | `wildfire-smoke-detection-v1.zip` | 13,518,198,597 | `27abdbe3d3703d9c6fa67bfde136088845726313e148644eb0e948f9a6211e13` | Training-ready for smoke detection | | |
| | Wildfire smoke segmentation | `wildfire-smoke-segmentation-v1.zip` | 464,864,048 | `23134190da8ef71b157764453f3d5575a339fe469878934c70e4972db33eee0e` | Dataset-ready; weak SAM masks and limited human masks | | |
| Total payload: 330,717,305,039 bytes. | |
| ## Integrity | |
| The accompanying `*.validation.json` and `*.zip.sha256` files record the checks applied | |
| before publication. All fifteen archives passed: | |
| - full ZIP CRC verification; | |
| - SHA-256 verification of every entry; | |
| - a single training root per archive; | |
| - path-traversal and duplicate-entry rejection; | |
| - cross-source exact-duplicate and split-leakage checks. | |
| The EO4Wildfires materialization contains exactly 31,730 scenes: 20,307 train, | |
| 5,077 validation, and 6,346 test. | |
| The cross-view registration bundle contains 390 verified pairs split into 288 train, | |
| 57 validation, and 45 test rows across 14 isolated spatial groups. It includes 264 | |
| AerialExtreMatch examples and 126 French rural or mountain ODM examples. The private | |
| double-validation lot and the failed RoMa quality probe remain excluded. | |
| The media-triage bundle derives image-level supervision only from verified source boxes and | |
| explicit negatives. It contains 155,044 images split into 107,449 train, 23,689 validation, | |
| and 23,906 test rows across 4,950 isolated groups. Touati and TaMduluza are not included: | |
| their published repository structure does not provide a reliable per-image class contract or | |
| a leakage-free video grouping. | |
| The orchestrator bundle contains 118 deterministic decision examples generated by the current | |
| FireWarning stage contracts and gate engine: 110 stage-gate cases and 8 consensus cases. Its two | |
| synthetic daily workflows cover a nominal path and a contradiction that invokes the final judge | |
| and abstains when raw evidence is unavailable. It contains no operational Die or Fontainebleau | |
| fact and never authorizes automatic publication. | |
| The FireSpread_MedEU bundle contains 103 event-isolated sequences and 316 usable cumulative | |
| burned-area targets. It does not claim observed active-front supervision. CrisisFACTS contributes | |
| 769 deduplicated facts over 26 event-days, while the structured IMSR bundle contains 15,982 | |
| incident-proxy sequences covering 88,208 incident-days; both remain out-of-domain for France until | |
| an independent French validation lot is approved. | |
| The Open Images subset contains 4,398 licensed images split into 3,310 train, 435 validation, and | |
| 653 test samples, with no exact pixel leakage. It retains human-verified boxes for ambulances, | |
| helicopters, and fixed-wing aircraft. Open Images has no boxable fire-engine class and does not | |
| identify aircraft by firefighting role, so this bundle is supplemental rather than production-ready. | |
| The TartanAir subset contains 34,742 aligned outdoor RGB, metric-depth, and camera-pose samples | |
| split by complete environment into 18,324 train, 5,838 validation, and 10,580 test rows. It is a | |
| synthetic geometry prior only: local NED poses are not geographic coordinates, and an independent | |
| real rural or mountain camera-pose test remains mandatory before operational use. | |
| The DIODE outdoor bundle contains 17,330 real outdoor RGB images aligned with metric depth and | |
| validity masks, split by complete scene into 13,678 train, 105 validation, and 3,547 test rows. | |
| Its 15 scenes have no split-group leakage. DIODE's official downloadable test partition is not | |
| available through the public archive used here, so the FireWarning splits are deterministically | |
| assigned from complete train and validation scenes and recorded in the source manifest. | |
| The Boreal detection subset contains 6,365 UAV images with 6,340 human smoke boxes and 256 | |
| documented empty-image negatives. Collection sites are isolated across 2,910 train, 2,217 | |
| validation, and 1,238 test rows. Fifty-two zero-byte labels outside the documented negative set | |
| are quarantined rather than silently converted into negative supervision. The segmentation subset | |
| contains 1,417 image-mask pairs: 40 human masks are strong labels and 1,377 SAM-generated masks | |
| remain weak supervision, recorded separately in every sample and split. | |
| ## Licenses and attribution | |
| This repository is a mixed-license collection. There is no single license covering every | |
| payload. Consult `TRAIN_BUNDLE.json` and the source metadata inside each ZIP before use or | |
| redistribution. Included source families declare licenses such as CC BY 4.0, | |
| CC BY-SA 4.0, CC0 1.0, Apache 2.0, MIT, and GPL 3.0, depending on the source. | |
| Operational evaluation packages for Die and Fontainebleau are not included. Some openly | |
| licensed third-party training candidates can still document those locations; their original | |
| source URL, author, and license are retained in the corresponding manifest. | |
| ## Publication policy | |
| These packages are training inputs, not operational fire reports. They must not be used as | |
| evidence of a current incident, and generated locations or perimeters still require the | |
| FireWarning deterministic gates and human validation before publication. | |
| ## Statut de publication FireViewer | |
| **Publication publique multi-licence.** Quinze ZIP64 sont publiés, mais seulement trois bundles sont déclarés training_ready. Les douze autres conservent leurs blocages documentés. Les contrats fire-pointing-lora-v1 et cross-view-localization-v1 ne doivent pas être exécutés tels quels ; le contrat de remplacement est mvp-a40-v2. | |
| ## Limites et sécurité | |
| > FireViewer est un projet de recherche et développement. Ce dataset n’est ni un service d’alerte, ni une source officielle, ni un outil de conduite des secours. Il ne doit pas être utilisé seul pour la sécurité des personnes, une décision d’intervention, une expertise d’assurance ou une preuve juridique. | |
| Les observations, reconstructions, annotations et simulations doivent être interprétées avec leur provenance, leur confiance et leurs limites. Une interpolation rétrospective n’est ni une observation directe ni une prévision. | |
| ## Droits | |
| Chaque source garde sa licence propre, notamment CC BY 4.0, CC BY-SA 4.0, CC0, Apache-2.0, MIT, GPL-3.0 ou des conditions spécifiques. Aucune licence unique ne couvre automatiquement les quinze bundles. | |
| ## Identité et contact | |
| FireViewer est un projet distinct maintenu par **Unicorn Who Dev**. Les espaces fireviewer sur GitHub et Hugging Face sont ses espaces de publication canoniques. | |
| Contact public, provenance, droits et demandes de retrait : [unicornwhodev@gmail.com](mailto:unicornwhodev@gmail.com). | |