fireviewer
/

litert-models

@@ -1,59 +1,99 @@
-# FireViewer · LiteRT on Android
-**Documented fire/smoke models with local adaptation and verifiable checkpoints.**
-Detection and pointing conversions for experimental Android integrations. Reviewed annotations can update heads or output adapters while visual backbones remain frozen. This package is not an operational fire-alert system.
-[Français](README.md) · **English**
-[Get started](docs/USAGE.en.md) · [Catalogue and evidence](docs/BENCHMARKS.en.md) · [Kotlin SDK](android/README.md) · [Continuous learning](docs/CONTINUOUS_LEARNING.en.md)
-| Delivered variants | Host training | Standalone Android |
-|---|---|---|
-| **6**, including 5 trainable | **5 checked conversions** | **YOLO11 M learning checked on emulator** |
-All six variants remain **unqualified inside Vision Dataset Studio**. The YOLO SDK pass covers neither other models nor application integration. Physical ARM, task accuracy and forgetting are unmeasured.
-## Find the component
-| Family | Files and scope |
-|---|---|
-| D-FINE M strict v1 | [Detection and candidate adaptation](models/fireviewer_dfine_m_strict_v1_learning/README.md) |
-| RT-DETRv2 R50 | [Detection and candidate adaptation](models/fireviewer_rtdetr_v2_r50_learning/README.md) |
-| YOLO11 M strict v1 | [Inference](models/fireviewer_yolo11m_strict_v1/README.md) · [Learning](models/fireviewer_yolo11m_strict_v1_learning/README.md) |
-| RF-DETR Medium 1.10 | [Detection and candidate adaptation](models/fireviewer_rfdetr_medium_v110_learning/README.md) |
-| DINOv3 pointing pilot v1 | [Segmentation, pointing, presence and abstention](models/fireviewer_dinov3_pointing_pilot_v1_learning/README.md) |
-The optional SegFormer-B2 source was unavailable; no SegFormer graph is delivered. Detection adapters cannot create candidates absent from the frozen detector.
-## Start with a controlled integration
-Choose a variant, read its contract and source notices, download a pinned revision and verify the artifact manifest. Apply its exact preprocessing, classes and coordinate transform. Test inference before training, use reviewed targets, keep the original and store checkpoints separately. Evaluate candidates before activation. [Full usage guide](docs/USAGE.en.md).
-## SDK and runtime
-The [Kotlin SDK](android/README.md) provides `infer`, `train`, `save` and `restore`. Checkpoints retain parameters, SGD momentum and step counters. Receipts verify model/label identity and file integrity; the integrating application decides which version is active.
-YOLO supports H/W multiples of 32. Other graphs resize to an internal grid. The historical SDK uses TensorFlow Lite + Select TF Ops **2.16.1**. Vision Dataset Studio's separate 16 KB Flex fix does not retroactively qualify this SDK.
-## Recorded qualification
-| Variante / Variant | Hôte / Host | SDK | Application / App |
-|---|---|---|---|
-| [D-FINE M · learning](models/fireviewer_dfine_m_strict_v1_learning/README.md) | PASS | — | Non qualifié / Not qualified |
-| [DINOv3 pointing · learning](models/fireviewer_dinov3_pointing_pilot_v1_learning/README.md) | PASS | — | Non qualifié / Not qualified |
-| [RF-DETR Medium · learning](models/fireviewer_rfdetr_medium_v110_learning/README.md) | PASS | — | Non qualifié / Not qualified |
-| [RT-DETRv2 R50 · learning](models/fireviewer_rtdetr_v2_r50_learning/README.md) | PASS | — | Non qualifié / Not qualified |
-| [YOLO11 M](models/fireviewer_yolo11m_strict_v1/README.md) | N/A | — | Non qualifié / Not qualified |
-| [YOLO11 M · learning](models/fireviewer_yolo11m_strict_v1_learning/README.md) | PASS | PASS | Non qualifié / Not qualified |
-![Historical conversion and SDK qualification matrix](docs/figures/qualification-matrix.png)
-Reports distinguish conversion parity, synthetic host learning, standalone Android SDK and application integration. [Full report](docs/BENCHMARKS.en.md) · [Delivery receipt](reports/fireviewer-completion.json) · [Conversion sources](conversion/README.md).
-## Rights and limitations
-Source models, datasets and components retain their own licences and restrictions. Check each model's rights and notices; **download access alone does not grant commercial-use rights**. This documentation changes no licence.
-Synthetic checks prove updates and persistence, not field accuracy, generalization or reliable emergency alerts. Select models using authorized representative data, human review and measurements on the intended device.
-<sub>Updated 23 September 2026 · Historical evidence preserved · No model weights changed.</sub>

+---
+library_name: tensorflow-lite
+license: other
+tags:
+- android
+- on-device-training
+- fire-detection
+---
+# FireViewer LiteRT Models for Android
+Fire/smoke detection and pointing conversions for experimental Android workflows:
+**6 delivered variants, including 5 with trainable heads or output adapters**.
+Reviewed annotations can update the exposed adaptation parameters while the
+visual backbones remain frozen.
+[Usage guide](docs/USAGE.en.md) · [Conversion and SDK evidence](docs/BENCHMARKS.en.md)
+· [Kotlin SDK](android/README.md) · [Checkpoint lifecycle](docs/CONTINUOUS_LEARNING.en.md).
+## Catalogue and recorded qualification
+| Variant | Synthetic host learning | Standalone Android SDK | Vision Dataset Studio |
+| --- | --- | --- | --- |
+| [D-FINE M strict v1, learning](models/fireviewer_dfine_m_strict_v1_learning/README.md) | PASS | Not qualified | Not qualified |
+| [RT-DETRv2 R50, learning](models/fireviewer_rtdetr_v2_r50_learning/README.md) | PASS | Not qualified | Not qualified |
+| [YOLO11 M strict v1, inference](models/fireviewer_yolo11m_strict_v1/README.md) | N/A | Not qualified | Not qualified |
+| [YOLO11 M strict v1, learning](models/fireviewer_yolo11m_strict_v1_learning/README.md) | PASS | PASS on emulator | Not qualified |
+| [RF-DETR Medium 1.10, learning](models/fireviewer_rfdetr_medium_v110_learning/README.md) | PASS | Not qualified | Not qualified |
+| [DINOv3 pointing pilot v1, learning](models/fireviewer_dinov3_pointing_pilot_v1_learning/README.md) | PASS | Not qualified | Not qualified |
+![Historical conversion and SDK qualification matrix](docs/figures/qualification-matrix.png)
+The retained [delivery receipt](reports/fireviewer-completion.json) covers five
+learning conversions. The YOLO inference-only graph is a sixth delivered variant.
+No SegFormer graph is delivered because the optional source was unavailable.
+Detection adapters cannot create proposals absent from the frozen detector.
+**All six variants remain unqualified inside Vision Dataset Studio.** The YOLO
+learning SDK pass does not qualify other models, application integration or a
+physical ARM phone. Task accuracy, generalization and forgetting are unmeasured.
+## Download and integrate
+Choose a variant and read its README, runtime contract, artifact manifest and
+source notices. Retrieve all matching files at the same immutable revision:
+```python
+from huggingface_hub import hf_hub_download
+contract = hf_hub_download(
+    repo_id="fireviewer/litert-models",
+    revision="bdcd483d5c5aadeff1b531a35b536f3530906e18",
+    filename="models/fireviewer_yolo11m_strict_v1_learning/runtime_contract.json",
+)
+```
+Apply the exact normalization, tensor layout, labels and coordinate transform.
+Check inference before learning, train only with reviewed targets, keep the
+original weights, and store checkpoints separately. Evaluate adapted candidates
+before activation. See the [complete usage guide](docs/USAGE.en.md).
+## SDK and runtime contract
+The Kotlin SDK exposes `infer`, `train`, `save` and `restore`. Checkpoints preserve
+parameters, SGD momentum and step counters; receipts check model/label identity
+and file integrity. The integrating application selects the active checkpoint.
+YOLO supports image dimensions that are multiples of 32. Other graphs resize to
+their internal grid. The historical SDK uses TensorFlow Lite and Select TF Ops
+**2.16.1**. Vision Dataset Studio's separate 16 KB Flex fix does not retroactively
+qualify this SDK. [Conversion sources](conversion/README.md).
+## Rights and responsible use
+Source weights, datasets and software retain their own licences and restrictions.
+Read each variant's notices; download access alone grants no commercial-use or
+redistribution permission. In particular, YOLO upstream obligations and the
+DINOv3 pilot's retained release policy remain applicable.
+Synthetic checks establish parameter updates and persistence, not field accuracy
+or reliable alerts. These models must not be used alone to confirm a fire,
+trigger an emergency alert, order an evacuation or direct emergency response.
+## Documentation and revision
+Card updated in English on **5 October 2026**, from the public repository at
+[`bdcd483d5c5aadeff1b531a35b536f3530906e18`](https://huggingface.co/fireviewer/litert-models/tree/bdcd483d5c5aadeff1b531a35b536f3530906e18) and its retained reports.
+This update checks documentation, repository metadata and small evidence files;
+it does not rerun training, inference, dataset payload verification or device qualification.
+Historical receipts keep their original dates, revisions and scope. Earlier README
+hashes in artifact manifests refer to those earlier releases.
+The association FIRE-VIEWER provides administrative and financial stewardship for
+resources under its control. Technical governance is maintainer-led. Upstream,
+source-specific and pre-association rights remain separate; repository placement
+does not transfer ownership or grant additional rights.
+[Project documentation](https://github.com/fireviewer/Fireviewer_doc/blob/main/docs/public/HUGGINGFACE.md)
+· [Institutional website](https://fire-viewer.fr). Contact: **contact@fire-viewer.fr**.

README.md CHANGED Viewed

@@ -7,68 +7,93 @@ tags:
 - fire-detection
 ---
-# FireViewer · LiteRT sur Android
-**Des modèles feu/fumée documentés, avec adaptation locale et checkpoints vérifiables.**
-Conversions de détection et de pointage pour des intégrations Android expérimentales. Les têtes ou sorties peuvent apprendre à partir d’annotations revues ; les réseaux visuels restent figés. Cette livraison ne constitue pas un système d’alerte incendie ou une validation opérationnelle.
-**Français** · [English](README.en.md)
-[Prise en main](docs/USAGE.md) · [Catalogue et preuves](docs/BENCHMARKS.md) · [SDK Kotlin](android/README.md) · [Apprentissage continu](docs/CONTINUOUS_LEARNING.md)
-| Variantes livrées | Apprentissage sur hôte | Android autonome |
-|---|---|---|
-| **6**, dont 5 entraînables | **5 conversions vérifiées** | **YOLO11 M learning vérifié sur émulateur** |
-Les six variantes restent **non qualifiées dans Vision Dataset Studio**. Le succès du SDK YOLO ne valide ni les autres modèles ni leur intégration dans l’application. Aucun téléphone ARM physique, score de détection ou test d’oubli n’est documenté.
-## Trouver le bon composant
-| Famille | Usage et fichiers |
-|---|---|
-| D-FINE M strict v1 | [Détection et adaptation des propositions](models/fireviewer_dfine_m_strict_v1_learning/README.md) |
-| RT-DETRv2 R50 | [Détection et adaptation des propositions](models/fireviewer_rtdetr_v2_r50_learning/README.md) |
-| YOLO11 M strict v1 | [Inférence](models/fireviewer_yolo11m_strict_v1/README.md) · [Variante entraînable](models/fireviewer_yolo11m_strict_v1_learning/README.md) |
-| RF-DETR Medium 1.10 | [Détection et adaptation des propositions](models/fireviewer_rfdetr_medium_v110_learning/README.md) |
-| DINOv3 pointing pilot v1 | [Segmentation, pointage, présence et abstention](models/fireviewer_dinov3_pointing_pilot_v1_learning/README.md) |
-La source optionnelle SegFormer-B2 n’était pas disponible : aucun modèle SegFormer n’est livré ici. Les adaptations de détection ne créent pas de propositions absentes du détecteur figé.
-## Démarrer sans ambiguïté
-1. Choisir une variante et lire son README, son contrat et ses notices.
-2. Télécharger une révision épinglée et vérifier le manifeste des artefacts.
-3. Appliquer exactement normalisation, disposition des tenseurs, classes et transformation des coordonnées.
-4. Tester l’inférence avant l’apprentissage ; fournir uniquement des annotations connues et revues.
-5. Conserver l’original, enregistrer les checkpoints séparément et évaluer un candidat avant activation.
-[Guide complet et exemples de téléchargement](docs/USAGE.md) · [Cycle des checkpoints](docs/CONTINUOUS_LEARNING.md).
-## SDK et architecture
-Le [SDK Kotlin](android/README.md) expose `infer`, `train`, `save` et `restore`. Ses générations conservent paramètres, moments SGD et compteur d’étapes ; les reçus contrôlent modèle, schéma de labels et intégrité des fichiers. L’application qui l’intègre reste responsable de la validation métier et du choix des poids actifs.
-YOLO accepte des dimensions H/W multiples de 32. Les autres modèles redimensionnent dans le graphe vers leur grille interne. Le SDK historique utilise TensorFlow Lite et Select TF Ops **2.16.1** ; le correctif Flex 16 Ko de Vision Dataset Studio est une livraison distincte, pas une qualification rétroactive de ce SDK.
-## Résultats documentés
-| Variante / Variant | Hôte / Host | SDK | Application / App |
-|---|---|---|---|
-| [D-FINE M · learning](models/fireviewer_dfine_m_strict_v1_learning/README.md) | PASS | — | Non qualifié / Not qualified |
-| [DINOv3 pointing · learning](models/fireviewer_dinov3_pointing_pilot_v1_learning/README.md) | PASS | — | Non qualifié / Not qualified |
-| [RF-DETR Medium · learning](models/fireviewer_rfdetr_medium_v110_learning/README.md) | PASS | — | Non qualifié / Not qualified |
-| [RT-DETRv2 R50 · learning](models/fireviewer_rtdetr_v2_r50_learning/README.md) | PASS | — | Non qualifié / Not qualified |
-| [YOLO11 M](models/fireviewer_yolo11m_strict_v1/README.md) | N/A | — | Non qualifié / Not qualified |
-| [YOLO11 M · learning](models/fireviewer_yolo11m_strict_v1_learning/README.md) | PASS | PASS | Non qualifié / Not qualified |
-![Matrice de qualification historique du SDK et des conversions](docs/figures/qualification-matrix.png)
-Les rapports séparent fidélité de conversion, apprentissage synthétique sur hôte, SDK Android et intégration applicative. [Rapport détaillé](docs/BENCHMARKS.md) · [Reçu de livraison](reports/fireviewer-completion.json) · [Sources de conversion](conversion/README.md).
-## Droits et utilisation responsable
-Les poids, jeux de données et composants amont gardent leurs licences et restrictions propres. Lire les fichiers de droits et notices de chaque modèle ; **la disponibilité au téléchargement n’accorde pas à elle seule un droit d’usage commercial**. Aucun droit n’est élargi par cette documentation.
-Les contrôles synthétiques montrent des mises à jour et leur persistance. Ils ne mesurent ni précision sur terrain, ni généralisation, ni capacité à déclencher une alerte fiable. La sélection finale exige un corpus autorisé représentatif, une revue humaine et des mesures sur l’appareil cible.
-<sub>Documentation actualisée le 23 septembre 2026 · Preuves historiques conservées · Aucun poids modifié.</sub>

 - fire-detection
 ---
+# FireViewer LiteRT Models for Android
+Fire/smoke detection and pointing conversions for experimental Android workflows:
+**6 delivered variants, including 5 with trainable heads or output adapters**.
+Reviewed annotations can update the exposed adaptation parameters while the
+visual backbones remain frozen.
+[Usage guide](docs/USAGE.en.md) · [Conversion and SDK evidence](docs/BENCHMARKS.en.md)
+· [Kotlin SDK](android/README.md) · [Checkpoint lifecycle](docs/CONTINUOUS_LEARNING.en.md).
+## Catalogue and recorded qualification
+| Variant | Synthetic host learning | Standalone Android SDK | Vision Dataset Studio |
+| --- | --- | --- | --- |
+| [D-FINE M strict v1, learning](models/fireviewer_dfine_m_strict_v1_learning/README.md) | PASS | Not qualified | Not qualified |
+| [RT-DETRv2 R50, learning](models/fireviewer_rtdetr_v2_r50_learning/README.md) | PASS | Not qualified | Not qualified |
+| [YOLO11 M strict v1, inference](models/fireviewer_yolo11m_strict_v1/README.md) | N/A | Not qualified | Not qualified |
+| [YOLO11 M strict v1, learning](models/fireviewer_yolo11m_strict_v1_learning/README.md) | PASS | PASS on emulator | Not qualified |
+| [RF-DETR Medium 1.10, learning](models/fireviewer_rfdetr_medium_v110_learning/README.md) | PASS | Not qualified | Not qualified |
+| [DINOv3 pointing pilot v1, learning](models/fireviewer_dinov3_pointing_pilot_v1_learning/README.md) | PASS | Not qualified | Not qualified |
+![Historical conversion and SDK qualification matrix](docs/figures/qualification-matrix.png)
+The retained [delivery receipt](reports/fireviewer-completion.json) covers five
+learning conversions. The YOLO inference-only graph is a sixth delivered variant.
+No SegFormer graph is delivered because the optional source was unavailable.
+Detection adapters cannot create proposals absent from the frozen detector.
+**All six variants remain unqualified inside Vision Dataset Studio.** The YOLO
+learning SDK pass does not qualify other models, application integration or a
+physical ARM phone. Task accuracy, generalization and forgetting are unmeasured.
+## Download and integrate
+Choose a variant and read its README, runtime contract, artifact manifest and
+source notices. Retrieve all matching files at the same immutable revision:
+```python
+from huggingface_hub import hf_hub_download
+contract = hf_hub_download(
+    repo_id="fireviewer/litert-models",
+    revision="bdcd483d5c5aadeff1b531a35b536f3530906e18",
+    filename="models/fireviewer_yolo11m_strict_v1_learning/runtime_contract.json",
+)
+```
+Apply the exact normalization, tensor layout, labels and coordinate transform.
+Check inference before learning, train only with reviewed targets, keep the
+original weights, and store checkpoints separately. Evaluate adapted candidates
+before activation. See the [complete usage guide](docs/USAGE.en.md).
+## SDK and runtime contract
+The Kotlin SDK exposes `infer`, `train`, `save` and `restore`. Checkpoints preserve
+parameters, SGD momentum and step counters; receipts check model/label identity
+and file integrity. The integrating application selects the active checkpoint.
+YOLO supports image dimensions that are multiples of 32. Other graphs resize to
+their internal grid. The historical SDK uses TensorFlow Lite and Select TF Ops
+**2.16.1**. Vision Dataset Studio's separate 16 KB Flex fix does not retroactively
+qualify this SDK. [Conversion sources](conversion/README.md).
+## Rights and responsible use
+Source weights, datasets and software retain their own licences and restrictions.
+Read each variant's notices; download access alone grants no commercial-use or
+redistribution permission. In particular, YOLO upstream obligations and the
+DINOv3 pilot's retained release policy remain applicable.
+Synthetic checks establish parameter updates and persistence, not field accuracy
+or reliable alerts. These models must not be used alone to confirm a fire,
+trigger an emergency alert, order an evacuation or direct emergency response.
+## Documentation and revision
+Card updated in English on **5 October 2026**, from the public repository at
+[`bdcd483d5c5aadeff1b531a35b536f3530906e18`](https://huggingface.co/fireviewer/litert-models/tree/bdcd483d5c5aadeff1b531a35b536f3530906e18) and its retained reports.
+This update checks documentation, repository metadata and small evidence files;
+it does not rerun training, inference, dataset payload verification or device qualification.
+Historical receipts keep their original dates, revisions and scope. Earlier README
+hashes in artifact manifests refer to those earlier releases.
+The association FIRE-VIEWER provides administrative and financial stewardship for
+resources under its control. Technical governance is maintainer-led. Upstream,
+source-specific and pre-association rights remain separate; repository placement
+does not transfer ownership or grant additional rights.
+[Project documentation](https://github.com/fireviewer/Fireviewer_doc/blob/main/docs/public/HUGGINGFACE.md)
+· [Institutional website](https://fire-viewer.fr). Contact: **contact@fire-viewer.fr**.