unicornwhodev commited on
Commit
a5bee07
·
verified ·
1 Parent(s): 9358081

Clean distribution files and correct FireViewer documentation

Browse files

Targeted cleanup authorized on 14 November 2026. Retain best weights, inference contracts, datasets, licenses and provenance. Remove identified resume/trainer/log artifacts and correct current documentation. No model training or evaluation performed.

docs/BENCHMARKS.en.md CHANGED
@@ -1,75 +1,75 @@
1
- # Benchmark evidence and qualification
2
-
3
- > **Documentation — 2026-09-23.** Catalogue actuel : 6 variantes, dont 5 entraînables. Les preuves ci-dessous gardent leur date et leur runtime. Aucun succès de Vision Dataset Studio n’est revendiqué pour ces six variantes. La correction Flex 16 Ko de rc5 est distincte de leur SDK historique. [Guide des versions](CONTINUOUS_LEARNING.md).
4
-
5
- > **Documentation update.** The six variants remain unqualified in Vision Dataset Studio. Historical host/SDK evidence below is unchanged. rc5 core tests do not qualify this SDK for 16 KB pages. [Version lifecycle](CONTINUOUS_LEARNING.en.md).
6
-
7
- [Français](BENCHMARKS.md) · [English](BENCHMARKS.en.md)
8
-
9
- **6 current conversion variants; 5 expose learning.** This report visualizes recorded checks; no new inference or training was run for this documentation update. The values are not a mobile performance leaderboard.
10
-
11
- **Source snapshot / Révision source :** `330e9097409042751988e9fa5994b51ac2b577bc` · evidence / preuves : 2026-09-20 · documentation : 2026-09-21.
12
-
13
- ![Qualification matrix](figures/qualification-matrix.png)
14
-
15
- **Five trainable conversions passed host checks, including TensorFlow Lite 2.16.1 restart and optimizer resume.** YOLO11 M additionally passed a standalone Android SDK test. All six variants still await execution through Vision Dataset Studio; the separate SDK result is not app integration. The optional SegFormer source was unavailable and has no delivered graph or benchmark.
16
-
17
- ## Catalogue / Catalogue
18
-
19
- | Conversion | Host train | Standalone SDK | App |
20
- |---|---|---|---|
21
- | [D-FINE M · learning](../models/fireviewer_dfine_m_strict_v1_learning/README.md) | PASS | PENDING | PENDING |
22
- | [DINOv3 pointing · learning](../models/fireviewer_dinov3_pointing_pilot_v1_learning/README.md) | PASS | PENDING | PENDING |
23
- | [RF-DETR Medium · learning](../models/fireviewer_rfdetr_medium_v110_learning/README.md) | PASS | PENDING | PENDING |
24
- | [RT-DETRv2 R50 · learning](../models/fireviewer_rtdetr_v2_r50_learning/README.md) | PASS | PENDING | PENDING |
25
- | [YOLO11 M](../models/fireviewer_yolo11m_strict_v1/README.md) | N/A | PENDING | PENDING |
26
- | [YOLO11 M · learning](../models/fireviewer_yolo11m_strict_v1_learning/README.md) | PASS | Reported PASS | PENDING |
27
-
28
- PASS is scoped to the indicated test. PENDING includes unexecuted or unfinished checks; it is not a failed test. N/A means there is no training signature.
29
-
30
- ## Host learning / Apprentissage de conversion sur hôte
31
-
32
- ![Host synthetic losses](figures/host-learning.png)
33
-
34
- | Conversion | Initial loss | Final reported loss | Restore | Resume | Trainable scope |
35
- |---|---|---|---|---|---|
36
- | fireviewer_dfine_m_strict_v1_learning | 2.38586688 | 1.08972955 | PASS | PASS | candidate_classification_and_box_adaptation_only |
37
- | fireviewer_dinov3_pointing_pilot_v1_learning | 4.22975874 | 3.05032301 | PASS | PASS | four_task_output_adaptation_only |
38
- | fireviewer_rfdetr_medium_v110_learning | 0.86913079 | 0.191484436 | PASS | PASS | candidate_classification_and_box_adaptation_only |
39
- | fireviewer_rtdetr_v2_r50_learning | 3.6428864 | 0.403154105 | PASS | PASS | candidate_classification_and_box_adaptation_only |
40
- | fireviewer_yolo11m_strict_v1_learning | 5.98196793 | 0.148539305 | PASS | PASS | candidate_classification_and_box_adaptation_only |
41
-
42
- Each pair comes from that model’s `training_validation.json`. Objectives, number of steps and input protocols differ; these are endpoints, not sampled learning curves. No common test dataset, confidence intervals, task accuracy or forgetting metrics were recorded. A zero loss on repeated synthetic examples does not imply perfect performance. All visual backbones are frozen.
43
-
44
- ## Standalone Android SDK / SDK Android autonome
45
-
46
- ![Standalone SDK results](figures/android-sdk.png)
47
-
48
- YOLO11 M learning: `ANDROID_FIRE_TRAINING_PASS`, 1 thread, 64×64 then 64×96 inputs, losses 7.3072906 and 7.0748734. These steps use different synthetic inputs and are not a fixed-sample learning curve. Exact restore, optimizer resume and incompatible-label-schema rejection pass. See the [complete log](../reports/android/fireviewer_yolo11m_strict_v1-final.log), including its startup/runtime warnings, and [runtime libraries](../reports/android/runtime-libraries.json). / Ces pertes concernent deux entrées différentes ; restauration, reprise et refus de schéma incompatible réussissent. Les avertissements du journal restent conservés.
49
-
50
- ## Conversion fidelity / Fidélité numérique
51
-
52
- ![Conversion parity](figures/conversion-parity.png)
53
-
54
- Each panel contains the source report’s absolute errors, in report order, across outputs/shapes/seeds. Output units and normalization differ, so values must not be used to rank architectures. The combined acceptance rule is `abs(actual-reference) <= atol + rtol * abs(reference)`; an absolute error above `atol` alone is not necessarily a failure. No task-quality conclusion follows from numerical parity.
55
-
56
- - [D-FINE M — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_dfine_m_strict_v1_learning/source_parity.json)
57
- - [DINOv3 · pointing — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_dinov3_pointing_pilot_v1_learning/source_parity.json)
58
- - [RF-DETR Medium — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_rfdetr_medium_v110_learning/source_parity.json)
59
- - [RT-DETRv2 R50 — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_rtdetr_v2_r50_learning/source_parity.json)
60
- - [YOLO11 M — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_yolo11m_strict_v1_learning/source_parity.json)
61
-
62
- D-FINE parity concerns the frozen component against ONNX, including protected decoder Softmax tensors and seeds 842/843/844; it is distinct from training-head validation. Do not remove the arena protection based on this documentation. / La parité D-FINE concerne le composant figé et les tenseurs protégés ; ne pas extrapoler à un graphe modifié.
63
-
64
- ## Reproduce / Reproduire
65
-
66
- ```bash
67
- python -m pip install matplotlib numpy
68
- python docs/evidence/plot_results.py docs/evidence/visual-data.json docs/figures
69
- ```
70
-
71
- The JSON contains exact plotted values, model hashes and source-file hashes. Figures are PNG + editable SVG, rendered with Matplotlib, not generated by an image model. The source snapshots are immutable; a new weight hash invalidates transfer of these results. This command only redraws figures and performs no model execution.
72
-
73
- [Usage and test protocol / Utilisation et protocole](USAGE.md) · [English usage guide](USAGE.en.md)
74
-
75
- Next required evidence: per-variant app inference and training where exposed; physical ARM device and runtime/delegate details; warm-up and repeated inference p50/p95 with failures retained; measured memory/thermal state; representative licensed held-out data with task-appropriate accuracy and forgetting analysis. No values are filled in for measurements that were not run.
 
1
+ # Benchmark evidence and qualification
2
+
3
+ > **Documentation — 2026-09-23.** Catalogue actuel : 6 variantes, dont 5 entraînables. Les preuves ci-dessous gardent leur date et leur runtime. Aucun succès de Vision Dataset Studio n’est revendiqué pour ces six variantes. La correction Flex 16 Ko de rc5 est distincte de leur SDK historique. [Guide des versions](CONTINUOUS_LEARNING.md).
4
+
5
+ > **Documentation update.** The six variants remain unqualified in Vision Dataset Studio. Historical host/SDK evidence below is unchanged. rc5 core tests do not qualify this SDK for 16 KB pages. [Version lifecycle](CONTINUOUS_LEARNING.en.md).
6
+
7
+ [Français](BENCHMARKS.md) · [English](BENCHMARKS.en.md)
8
+
9
+ **6 current conversion variants; 5 expose learning.** This report visualizes recorded checks; no new inference or training was run for this documentation update. The values are not a mobile performance leaderboard.
10
+
11
+ **Source snapshot / Révision source :** `330e9097409042751988e9fa5994b51ac2b577bc` · evidence / preuves : 2026-09-20 · documentation : 2026-09-21.
12
+
13
+ ![Qualification matrix](figures/qualification-matrix.png)
14
+
15
+ **Five trainable conversions passed host checks, including TensorFlow Lite 2.16.1 restart and optimizer resume.** YOLO11 M additionally passed a standalone Android SDK test. All six variants still await execution through Vision Dataset Studio; the separate SDK result is not app integration. The optional SegFormer source was unavailable and has no delivered graph or benchmark.
16
+
17
+ ## Catalogue / Catalogue
18
+
19
+ | Conversion | Host train | Standalone SDK | App |
20
+ |---|---|---|---|
21
+ | [D-FINE M · learning](../models/fireviewer_dfine_m_strict_v1_learning/README.md) | PASS | PENDING | PENDING |
22
+ | [DINOv3 pointing · learning](../models/fireviewer_dinov3_pointing_pilot_v1_learning/README.md) | PASS | PENDING | PENDING |
23
+ | [RF-DETR Medium · learning](../models/fireviewer_rfdetr_medium_v110_learning/README.md) | PASS | PENDING | PENDING |
24
+ | [RT-DETRv2 R50 · learning](../models/fireviewer_rtdetr_v2_r50_learning/README.md) | PASS | PENDING | PENDING |
25
+ | [YOLO11 M](../models/fireviewer_yolo11m_strict_v1/README.md) | N/A | PENDING | PENDING |
26
+ | [YOLO11 M · learning](../models/fireviewer_yolo11m_strict_v1_learning/README.md) | PASS | Reported PASS | PENDING |
27
+
28
+ PASS is scoped to the indicated test. PENDING includes unexecuted or unfinished checks; it is not a failed test. N/A means there is no training signature.
29
+
30
+ ## Host learning / Apprentissage de conversion sur hôte
31
+
32
+ ![Host synthetic losses](figures/host-learning.png)
33
+
34
+ | Conversion | Initial loss | Final reported loss | Restore | Resume | Trainable scope |
35
+ |---|---|---|---|---|---|
36
+ | fireviewer_dfine_m_strict_v1_learning | 2.38586688 | 1.08972955 | PASS | PASS | candidate_classification_and_box_adaptation_only |
37
+ | fireviewer_dinov3_pointing_pilot_v1_learning | 4.22975874 | 3.05032301 | PASS | PASS | four_task_output_adaptation_only |
38
+ | fireviewer_rfdetr_medium_v110_learning | 0.86913079 | 0.191484436 | PASS | PASS | candidate_classification_and_box_adaptation_only |
39
+ | fireviewer_rtdetr_v2_r50_learning | 3.6428864 | 0.403154105 | PASS | PASS | candidate_classification_and_box_adaptation_only |
40
+ | fireviewer_yolo11m_strict_v1_learning | 5.98196793 | 0.148539305 | PASS | PASS | candidate_classification_and_box_adaptation_only |
41
+
42
+ Each pair comes from that model’s `training_validation.json`. Objectives, number of steps and input protocols differ; these are endpoints, not sampled learning curves. No common test dataset, confidence intervals, task accuracy or forgetting metrics were recorded. A zero loss on repeated synthetic examples does not imply perfect performance. All visual backbones are frozen.
43
+
44
+ ## Standalone Android SDK / SDK Android autonome
45
+
46
+ ![Standalone SDK results](figures/android-sdk.png)
47
+
48
+ YOLO11 M learning: `ANDROID_FIRE_TRAINING_PASS`, 1 thread, 64×64 then 64×96 inputs, losses 7.3072906 and 7.0748734. These steps use different synthetic inputs and are not a fixed-sample learning curve. Exact restore, optimizer resume and incompatible-label-schema rejection pass. See the [archived log](https://huggingface.co/fireviewer/litert-models/blob/9358081e49af8c608c44baa5016ca4910ec32180/reports/android/fireviewer_yolo11m_strict_v1-final.log), including its startup/runtime warnings, and [runtime libraries](../reports/android/runtime-libraries.json). / Ces pertes concernent deux entrées différentes ; restauration, reprise et refus de schéma incompatible réussissent. Le journal a été retiré de la branche courante ; la révision historique liée conserve les avertissements dex2oat 32/64 bits et Restore/use_inter_op_parallelism.
49
+
50
+ ## Conversion fidelity / Fidélité numérique
51
+
52
+ ![Conversion parity](figures/conversion-parity.png)
53
+
54
+ Each panel contains the source report’s absolute errors, in report order, across outputs/shapes/seeds. Output units and normalization differ, so values must not be used to rank architectures. The combined acceptance rule is `abs(actual-reference) <= atol + rtol * abs(reference)`; an absolute error above `atol` alone is not necessarily a failure. No task-quality conclusion follows from numerical parity.
55
+
56
+ - [D-FINE M — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_dfine_m_strict_v1_learning/source_parity.json)
57
+ - [DINOv3 · pointing — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_dinov3_pointing_pilot_v1_learning/source_parity.json)
58
+ - [RF-DETR Medium — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_rfdetr_medium_v110_learning/source_parity.json)
59
+ - [RT-DETRv2 R50 — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_rtdetr_v2_r50_learning/source_parity.json)
60
+ - [YOLO11 M — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_yolo11m_strict_v1_learning/source_parity.json)
61
+
62
+ D-FINE parity concerns the frozen component against ONNX, including protected decoder Softmax tensors and seeds 842/843/844; it is distinct from training-head validation. Do not remove the arena protection based on this documentation. / La parité D-FINE concerne le composant figé et les tenseurs protégés ; ne pas extrapoler à un graphe modifié.
63
+
64
+ ## Reproduce / Reproduire
65
+
66
+ ```bash
67
+ python -m pip install matplotlib numpy
68
+ python docs/evidence/plot_results.py docs/evidence/visual-data.json docs/figures
69
+ ```
70
+
71
+ The JSON contains exact plotted values, model hashes and source-file hashes. Figures are PNG + editable SVG, rendered with Matplotlib, not generated by an image model. The source snapshots are immutable; a new weight hash invalidates transfer of these results. This command only redraws figures and performs no model execution.
72
+
73
+ [Usage and test protocol / Utilisation et protocole](USAGE.md) · [English usage guide](USAGE.en.md)
74
+
75
+ Next required evidence: per-variant app inference and training where exposed; physical ARM device and runtime/delegate details; warm-up and repeated inference p50/p95 with failures retained; measured memory/thermal state; representative licensed held-out data with task-appropriate accuracy and forgetting analysis. No values are filled in for measurements that were not run.
docs/BENCHMARKS.md CHANGED
@@ -1,75 +1,75 @@
1
- # Résultats de contrôle et qualification des modèles
2
-
3
- > **Documentation — 2026-09-23.** Catalogue actuel : 6 variantes, dont 5 entraînables. Les preuves ci-dessous gardent leur date et leur runtime. Aucun succès de Vision Dataset Studio n’est revendiqué pour ces six variantes. La correction Flex 16 Ko de rc5 est distincte de leur SDK historique. [Guide des versions](CONTINUOUS_LEARNING.md).
4
-
5
- > **Documentation update.** The six variants remain unqualified in Vision Dataset Studio. Historical host/SDK evidence below is unchanged. rc5 core tests do not qualify this SDK for 16 KB pages. [Version lifecycle](CONTINUOUS_LEARNING.en.md).
6
-
7
- [Français](BENCHMARKS.md) · [English](BENCHMARKS.en.md)
8
-
9
- **6 variantes de conversion actives ; 5 exposent l’apprentissage.** Ce rapport illustre les contrôles enregistrés. Aucun nouveau calcul d’inférence ou d’apprentissage n’a été lancé pour cette actualisation. Ces valeurs ne constituent pas un classement de performances mobiles.
10
-
11
- **Source snapshot / Révision source :** `330e9097409042751988e9fa5994b51ac2b577bc` · evidence / preuves : 2026-09-20 · documentation : 2026-09-21.
12
-
13
- ![Qualification matrix](figures/qualification-matrix.png)
14
-
15
- **Cinq conversions entraînables ont passé les contrôles hôte, dont restauration et reprise sous TensorFlow Lite 2.16.1.** YOLO11 M a aussi passé un test SDK Android autonome. Les six variantes attendent leur exécution dans Vision Dataset Studio ; le SDK ne valide pas cette intégration. La source optionnelle SegFormer était indisponible : aucun graphe ni benchmark livré.
16
-
17
- ## Catalogue / Catalogue
18
-
19
- | Conversion | Host train | Standalone SDK | App |
20
- |---|---|---|---|
21
- | [D-FINE M · learning](../models/fireviewer_dfine_m_strict_v1_learning/README.md) | PASS | PENDING | PENDING |
22
- | [DINOv3 pointing · learning](../models/fireviewer_dinov3_pointing_pilot_v1_learning/README.md) | PASS | PENDING | PENDING |
23
- | [RF-DETR Medium · learning](../models/fireviewer_rfdetr_medium_v110_learning/README.md) | PASS | PENDING | PENDING |
24
- | [RT-DETRv2 R50 · learning](../models/fireviewer_rtdetr_v2_r50_learning/README.md) | PASS | PENDING | PENDING |
25
- | [YOLO11 M](../models/fireviewer_yolo11m_strict_v1/README.md) | N/A | PENDING | PENDING |
26
- | [YOLO11 M · learning](../models/fireviewer_yolo11m_strict_v1_learning/README.md) | PASS | Reported PASS | PENDING |
27
-
28
- PASS vaut uniquement pour le contrôle indiqué. PENDING couvre un contrôle non exécuté ou inachevé, sans signifier un échec. N/A indique l’absence de signature d’apprentissage.
29
-
30
- ## Host learning / Apprentissage de conversion sur hôte
31
-
32
- ![Host synthetic losses](figures/host-learning.png)
33
-
34
- | Conversion | Initial loss | Final reported loss | Restore | Resume | Trainable scope |
35
- |---|---|---|---|---|---|
36
- | fireviewer_dfine_m_strict_v1_learning | 2.38586688 | 1.08972955 | PASS | PASS | candidate_classification_and_box_adaptation_only |
37
- | fireviewer_dinov3_pointing_pilot_v1_learning | 4.22975874 | 3.05032301 | PASS | PASS | four_task_output_adaptation_only |
38
- | fireviewer_rfdetr_medium_v110_learning | 0.86913079 | 0.191484436 | PASS | PASS | candidate_classification_and_box_adaptation_only |
39
- | fireviewer_rtdetr_v2_r50_learning | 3.6428864 | 0.403154105 | PASS | PASS | candidate_classification_and_box_adaptation_only |
40
- | fireviewer_yolo11m_strict_v1_learning | 5.98196793 | 0.148539305 | PASS | PASS | candidate_classification_and_box_adaptation_only |
41
-
42
- Chaque paire provient du fichier `training_validation.json` du modèle. Objectifs, nombres d’étapes et protocoles d’entrée diffèrent : ce sont des valeurs de début/fin, pas des courbes échantillonnées. Aucun corpus commun, intervalle de confiance, score métier ou mesure d’oubli n’a été enregistré. Une perte nulle sur des exemples synthétiques répétés ne signifie pas une précision parfaite. Tous les encodeurs visuels restent figés.
43
-
44
- ## Standalone Android SDK / SDK Android autonome
45
-
46
- ![Standalone SDK results](figures/android-sdk.png)
47
-
48
- YOLO11 M learning: `ANDROID_FIRE_TRAINING_PASS`, 1 thread, 64×64 then 64×96 inputs, losses 7.3072906 and 7.0748734. These steps use different synthetic inputs and are not a fixed-sample learning curve. Exact restore, optimizer resume and incompatible-label-schema rejection pass. See the [complete log](../reports/android/fireviewer_yolo11m_strict_v1-final.log), including its startup/runtime warnings, and [runtime libraries](../reports/android/runtime-libraries.json). / Ces pertes concernent deux entrées différentes ; restauration, reprise et refus de schéma incompatible réussissent. Les avertissements du journal restent conservés.
49
-
50
- ## Conversion fidelity / Fidélité numérique
51
-
52
- ![Conversion parity](figures/conversion-parity.png)
53
-
54
- Chaque panneau reprend les erreurs absolues du rapport, dans leur ordre, selon les sorties/formes/graines. Les unités et normalisations diffèrent : aucun classement des architectures n’en découle. La règle combine `abs(obtenu-référence) <= atol + rtol * abs(référence)` ; dépasser `atol` seul ne suffit pas à conclure à un échec. La fidélité numérique ne démontre pas une précision métier.
55
-
56
- - [D-FINE M — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_dfine_m_strict_v1_learning/source_parity.json)
57
- - [DINOv3 · pointing — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_dinov3_pointing_pilot_v1_learning/source_parity.json)
58
- - [RF-DETR Medium — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_rfdetr_medium_v110_learning/source_parity.json)
59
- - [RT-DETRv2 R50 — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_rtdetr_v2_r50_learning/source_parity.json)
60
- - [YOLO11 M — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_yolo11m_strict_v1_learning/source_parity.json)
61
-
62
- D-FINE parity concerns the frozen component against ONNX, including protected decoder Softmax tensors and seeds 842/843/844; it is distinct from training-head validation. Do not remove the arena protection based on this documentation. / La parité D-FINE concerne le composant figé et les tenseurs protégés ; ne pas extrapoler à un graphe modifié.
63
-
64
- ## Reproduce / Reproduire
65
-
66
- ```bash
67
- python -m pip install matplotlib numpy
68
- python docs/evidence/plot_results.py docs/evidence/visual-data.json docs/figures
69
- ```
70
-
71
- Le JSON contient les valeurs exactes, empreintes des modèles et fichiers sources. Les figures PNG + SVG sont tracées avec Matplotlib, pas générées par un modèle d’image. Les révisions sont immuables ; un nouveau hash de poids impose de nouveaux tests. Cette commande retrace uniquement les figures, sans exécuter de modèle.
72
-
73
- [Usage and test protocol / Utilisation et protocole](USAGE.md) · [English usage guide](USAGE.en.md)
74
-
75
- À compléter : exécution de chaque variante dans l’application ; téléphone ARM avec runtime/délégué précisés ; préchauffage et répétitions p50/p95 avec échecs conservés ; mémoire et chauffe ; corpus représentatif autorisé, précision métier et oubli mesurés sur un jeu réservé. Aucune valeur n’est renseignée pour une mesure non exécutée.
 
1
+ # Résultats de contrôle et qualification des modèles
2
+
3
+ > **Documentation — 2026-09-23.** Catalogue actuel : 6 variantes, dont 5 entraînables. Les preuves ci-dessous gardent leur date et leur runtime. Aucun succès de Vision Dataset Studio n’est revendiqué pour ces six variantes. La correction Flex 16 Ko de rc5 est distincte de leur SDK historique. [Guide des versions](CONTINUOUS_LEARNING.md).
4
+
5
+ > **Documentation update.** The six variants remain unqualified in Vision Dataset Studio. Historical host/SDK evidence below is unchanged. rc5 core tests do not qualify this SDK for 16 KB pages. [Version lifecycle](CONTINUOUS_LEARNING.en.md).
6
+
7
+ [Français](BENCHMARKS.md) · [English](BENCHMARKS.en.md)
8
+
9
+ **6 variantes de conversion actives ; 5 exposent l’apprentissage.** Ce rapport illustre les contrôles enregistrés. Aucun nouveau calcul d’inférence ou d’apprentissage n’a été lancé pour cette actualisation. Ces valeurs ne constituent pas un classement de performances mobiles.
10
+
11
+ **Source snapshot / Révision source :** `330e9097409042751988e9fa5994b51ac2b577bc` · evidence / preuves : 2026-09-20 · documentation : 2026-09-21.
12
+
13
+ ![Qualification matrix](figures/qualification-matrix.png)
14
+
15
+ **Cinq conversions entraînables ont passé les contrôles hôte, dont restauration et reprise sous TensorFlow Lite 2.16.1.** YOLO11 M a aussi passé un test SDK Android autonome. Les six variantes attendent leur exécution dans Vision Dataset Studio ; le SDK ne valide pas cette intégration. La source optionnelle SegFormer était indisponible : aucun graphe ni benchmark livré.
16
+
17
+ ## Catalogue / Catalogue
18
+
19
+ | Conversion | Host train | Standalone SDK | App |
20
+ |---|---|---|---|
21
+ | [D-FINE M · learning](../models/fireviewer_dfine_m_strict_v1_learning/README.md) | PASS | PENDING | PENDING |
22
+ | [DINOv3 pointing · learning](../models/fireviewer_dinov3_pointing_pilot_v1_learning/README.md) | PASS | PENDING | PENDING |
23
+ | [RF-DETR Medium · learning](../models/fireviewer_rfdetr_medium_v110_learning/README.md) | PASS | PENDING | PENDING |
24
+ | [RT-DETRv2 R50 · learning](../models/fireviewer_rtdetr_v2_r50_learning/README.md) | PASS | PENDING | PENDING |
25
+ | [YOLO11 M](../models/fireviewer_yolo11m_strict_v1/README.md) | N/A | PENDING | PENDING |
26
+ | [YOLO11 M · learning](../models/fireviewer_yolo11m_strict_v1_learning/README.md) | PASS | Reported PASS | PENDING |
27
+
28
+ PASS vaut uniquement pour le contrôle indiqué. PENDING couvre un contrôle non exécuté ou inachevé, sans signifier un échec. N/A indique l’absence de signature d’apprentissage.
29
+
30
+ ## Host learning / Apprentissage de conversion sur hôte
31
+
32
+ ![Host synthetic losses](figures/host-learning.png)
33
+
34
+ | Conversion | Initial loss | Final reported loss | Restore | Resume | Trainable scope |
35
+ |---|---|---|---|---|---|
36
+ | fireviewer_dfine_m_strict_v1_learning | 2.38586688 | 1.08972955 | PASS | PASS | candidate_classification_and_box_adaptation_only |
37
+ | fireviewer_dinov3_pointing_pilot_v1_learning | 4.22975874 | 3.05032301 | PASS | PASS | four_task_output_adaptation_only |
38
+ | fireviewer_rfdetr_medium_v110_learning | 0.86913079 | 0.191484436 | PASS | PASS | candidate_classification_and_box_adaptation_only |
39
+ | fireviewer_rtdetr_v2_r50_learning | 3.6428864 | 0.403154105 | PASS | PASS | candidate_classification_and_box_adaptation_only |
40
+ | fireviewer_yolo11m_strict_v1_learning | 5.98196793 | 0.148539305 | PASS | PASS | candidate_classification_and_box_adaptation_only |
41
+
42
+ Chaque paire provient du fichier `training_validation.json` du modèle. Objectifs, nombres d’étapes et protocoles d’entrée diffèrent : ce sont des valeurs de début/fin, pas des courbes échantillonnées. Aucun corpus commun, intervalle de confiance, score métier ou mesure d’oubli n’a été enregistré. Une perte nulle sur des exemples synthétiques répétés ne signifie pas une précision parfaite. Tous les encodeurs visuels restent figés.
43
+
44
+ ## Standalone Android SDK / SDK Android autonome
45
+
46
+ ![Standalone SDK results](figures/android-sdk.png)
47
+
48
+ YOLO11 M learning: `ANDROID_FIRE_TRAINING_PASS`, 1 thread, 64×64 then 64×96 inputs, losses 7.3072906 and 7.0748734. These steps use different synthetic inputs and are not a fixed-sample learning curve. Exact restore, optimizer resume and incompatible-label-schema rejection pass. See the [archived log](https://huggingface.co/fireviewer/litert-models/blob/9358081e49af8c608c44baa5016ca4910ec32180/reports/android/fireviewer_yolo11m_strict_v1-final.log), including its startup/runtime warnings, and [runtime libraries](../reports/android/runtime-libraries.json). / Ces pertes concernent deux entrées différentes ; restauration, reprise et refus de schéma incompatible réussissent. Le journal a été retiré de la branche courante ; la révision historique liée conserve les avertissements dex2oat 32/64 bits et Restore/use_inter_op_parallelism.
49
+
50
+ ## Conversion fidelity / Fidélité numérique
51
+
52
+ ![Conversion parity](figures/conversion-parity.png)
53
+
54
+ Chaque panneau reprend les erreurs absolues du rapport, dans leur ordre, selon les sorties/formes/graines. Les unités et normalisations diffèrent : aucun classement des architectures n’en découle. La règle combine `abs(obtenu-référence) <= atol + rtol * abs(référence)` ; dépasser `atol` seul ne suffit pas à conclure à un échec. La fidélité numérique ne démontre pas une précision métier.
55
+
56
+ - [D-FINE M — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_dfine_m_strict_v1_learning/source_parity.json)
57
+ - [DINOv3 · pointing — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_dinov3_pointing_pilot_v1_learning/source_parity.json)
58
+ - [RF-DETR Medium — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_rfdetr_medium_v110_learning/source_parity.json)
59
+ - [RT-DETRv2 R50 — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_rtdetr_v2_r50_learning/source_parity.json)
60
+ - [YOLO11 M — original report](https://huggingface.co/fireviewer/litert-models/blob/330e9097409042751988e9fa5994b51ac2b577bc/models/fireviewer_yolo11m_strict_v1_learning/source_parity.json)
61
+
62
+ D-FINE parity concerns the frozen component against ONNX, including protected decoder Softmax tensors and seeds 842/843/844; it is distinct from training-head validation. Do not remove the arena protection based on this documentation. / La parité D-FINE concerne le composant figé et les tenseurs protégés ; ne pas extrapoler à un graphe modifié.
63
+
64
+ ## Reproduce / Reproduire
65
+
66
+ ```bash
67
+ python -m pip install matplotlib numpy
68
+ python docs/evidence/plot_results.py docs/evidence/visual-data.json docs/figures
69
+ ```
70
+
71
+ Le JSON contient les valeurs exactes, empreintes des modèles et fichiers sources. Les figures PNG + SVG sont tracées avec Matplotlib, pas générées par un modèle d’image. Les révisions sont immuables ; un nouveau hash de poids impose de nouveaux tests. Cette commande retrace uniquement les figures, sans exécuter de modèle.
72
+
73
+ [Usage and test protocol / Utilisation et protocole](USAGE.md) · [English usage guide](USAGE.en.md)
74
+
75
+ À compléter : exécution de chaque variante dans l’application ; téléphone ARM avec runtime/délégué précisés ; préchauffage et répétitions p50/p95 avec échecs conservés ; mémoire et chauffe ; corpus représentatif autorisé, précision métier et oubli mesurés sur un jeu réservé. Aucune valeur n’est renseignée pour une mesure non exécutée.
reports/android/fire-build-final.log DELETED
@@ -1,3 +0,0 @@
1
- warning: [options] system modules path not set in conjunction with -source 11
2
- 1 warning
3
- /workspace/qa/vds-model-training/training-check.jar
 
 
 
 
reports/android/fireviewer_yolo11m_strict_v1-final.log DELETED
@@ -1,25 +0,0 @@
1
- CANNOT LINK EXECUTABLE "/system/bin/dex2oat": "/system/lib64/libart-compiler.so" is 64-bit instead of 32-bit
2
- FIRE_CHECK opening yolo-final.tflite
3
- INFO: Initialized TensorFlow Lite runtime.
4
- INFO: Created TensorFlow Lite delegate for select TF ops.
5
- native : core/platform/cpu_feature_guard.cc:210 This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
6
- To enable the following instructions: SSE3 SSE4.1 SSE4.2, in other operations, rebuild TensorFlow with the appropriate compiler flags.
7
- INFO: TfLiteFlexDelegate delegate: 0 nodes delegated out of 1270 nodes with 0 partitions.
8
-
9
- VERBOSE: Replacing 9 out of 29 node(s) with delegate (TfLiteFlexDelegate) node, yielding 3 partitions for the whole graph.
10
- native : core/framework/node_def_util.cc:676 NodeDef mentions attribute use_inter_op_parallelism which is not in the op definition: Op<name=Restore; signature=file_pattern:string, tensor_name:string -> tensor:dt; attr=dt:type; attr=preferred_shard:int,default=-1; is_stateful=true> This may be expected if your graph generating binary is newer than this binary. Unknown attributes will be ignored. NodeDef: {{node Restore}}
11
- VERBOSE: Replacing 1 out of 20 node(s) with delegate (TfLiteFlexDelegate) node, yielding 2 partitions for the whole graph.
12
- VERBOSE: Replacing 9 out of 1537 node(s) with delegate (TfLiteFlexDelegate) node, yielding 5 partitions for the whole graph.
13
- FIRE_CHECK opened
14
- FIRE_CHECK infer 504
15
- FIRE_CHECK saved loss=7.3072906
16
- FIRE_CHECK continued loss=7.0748734
17
- VERBOSE: Replacing 9 out of 29 node(s) with delegate (TfLiteFlexDelegate) node, yielding 3 partitions for the whole graph.
18
- native : core/framework/node_def_util.cc:676 NodeDef mentions attribute use_inter_op_parallelism which is not in the op definition: Op<name=Restore; signature=file_pattern:string, tensor_name:string -> tensor:dt; attr=dt:type; attr=preferred_shard:int,default=-1; is_stateful=true> This may be expected if your graph generating binary is newer than this binary. Unknown attributes will be ignored. NodeDef: {{node Restore}}
19
- VERBOSE: Replacing 1 out of 20 node(s) with delegate (TfLiteFlexDelegate) node, yielding 2 partitions for the whole graph.
20
- VERBOSE: Replacing 9 out of 1537 node(s) with delegate (TfLiteFlexDelegate) node, yielding 5 partitions for the whole graph.
21
- VERBOSE: Replacing 9 out of 29 node(s) with delegate (TfLiteFlexDelegate) node, yielding 3 partitions for the whole graph.
22
- native : core/framework/node_def_util.cc:676 NodeDef mentions attribute use_inter_op_parallelism which is not in the op definition: Op<name=Restore; signature=file_pattern:string, tensor_name:string -> tensor:dt; attr=dt:type; attr=preferred_shard:int,default=-1; is_stateful=true> This may be expected if your graph generating binary is newer than this binary. Unknown attributes will be ignored. NodeDef: {{node Restore}}
23
- VERBOSE: Replacing 1 out of 20 node(s) with delegate (TfLiteFlexDelegate) node, yielding 2 partitions for the whole graph.
24
- VERBOSE: Replacing 9 out of 1537 node(s) with delegate (TfLiteFlexDelegate) node, yielding 5 partitions for the whole graph.
25
- ANDROID_FIRE_TRAINING_PASS {"runtime":"2.16.1","threads":1,"mixed_shapes":[[64,64],[64,96]],"train":true,"restore_exact":true,"optimizer_resume_exact":true,"incompatible_schema_rejected":true,"first_loss":7.3072906,"second_loss":7.0748734}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
reports/android/fireviewer_yolo11m_strict_v1.log DELETED
@@ -1,25 +0,0 @@
1
- CANNOT LINK EXECUTABLE "/system/bin/dex2oat": "/system/lib64/libart-compiler.so" is 64-bit instead of 32-bit
2
- FIRE_CHECK opening yolo.tflite
3
- INFO: Initialized TensorFlow Lite runtime.
4
- INFO: Created TensorFlow Lite delegate for select TF ops.
5
- native : core/platform/cpu_feature_guard.cc:210 This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
6
- To enable the following instructions: SSE3 SSE4.1 SSE4.2, in other operations, rebuild TensorFlow with the appropriate compiler flags.
7
- INFO: TfLiteFlexDelegate delegate: 0 nodes delegated out of 1270 nodes with 0 partitions.
8
-
9
- VERBOSE: Replacing 9 out of 29 node(s) with delegate (TfLiteFlexDelegate) node, yielding 3 partitions for the whole graph.
10
- native : core/framework/node_def_util.cc:676 NodeDef mentions attribute use_inter_op_parallelism which is not in the op definition: Op<name=Restore; signature=file_pattern:string, tensor_name:string -> tensor:dt; attr=dt:type; attr=preferred_shard:int,default=-1; is_stateful=true> This may be expected if your graph generating binary is newer than this binary. Unknown attributes will be ignored. NodeDef: {{node Restore}}
11
- VERBOSE: Replacing 1 out of 20 node(s) with delegate (TfLiteFlexDelegate) node, yielding 2 partitions for the whole graph.
12
- VERBOSE: Replacing 9 out of 1532 node(s) with delegate (TfLiteFlexDelegate) node, yielding 5 partitions for the whole graph.
13
- FIRE_CHECK opened
14
- FIRE_CHECK infer 2016
15
- FIRE_CHECK saved loss=7.079305
16
- FIRE_CHECK continued loss=7.035917
17
- VERBOSE: Replacing 9 out of 29 node(s) with delegate (TfLiteFlexDelegate) node, yielding 3 partitions for the whole graph.
18
- native : core/framework/node_def_util.cc:676 NodeDef mentions attribute use_inter_op_parallelism which is not in the op definition: Op<name=Restore; signature=file_pattern:string, tensor_name:string -> tensor:dt; attr=dt:type; attr=preferred_shard:int,default=-1; is_stateful=true> This may be expected if your graph generating binary is newer than this binary. Unknown attributes will be ignored. NodeDef: {{node Restore}}
19
- VERBOSE: Replacing 1 out of 20 node(s) with delegate (TfLiteFlexDelegate) node, yielding 2 partitions for the whole graph.
20
- VERBOSE: Replacing 9 out of 1532 node(s) with delegate (TfLiteFlexDelegate) node, yielding 5 partitions for the whole graph.
21
- VERBOSE: Replacing 9 out of 29 node(s) with delegate (TfLiteFlexDelegate) node, yielding 3 partitions for the whole graph.
22
- native : core/framework/node_def_util.cc:676 NodeDef mentions attribute use_inter_op_parallelism which is not in the op definition: Op<name=Restore; signature=file_pattern:string, tensor_name:string -> tensor:dt; attr=dt:type; attr=preferred_shard:int,default=-1; is_stateful=true> This may be expected if your graph generating binary is newer than this binary. Unknown attributes will be ignored. NodeDef: {{node Restore}}
23
- VERBOSE: Replacing 1 out of 20 node(s) with delegate (TfLiteFlexDelegate) node, yielding 2 partitions for the whole graph.
24
- VERBOSE: Replacing 9 out of 1532 node(s) with delegate (TfLiteFlexDelegate) node, yielding 5 partitions for the whole graph.
25
- ANDROID_FIRE_TRAINING_PASS {"runtime":"2.16.1","mixed_shapes":[[128,128],[128,192]],"train":true,"restore_exact":true,"optimizer_resume_exact":true,"incompatible_schema_rejected":true,"first_loss":7.079305,"second_loss":7.035917}