CoolFace
Modelpublic

Charlbi/Lite_rt_prepared_for_android_dataset_builder

sourceHugging Faceupdated 1d agoView on Hugging Face
0likes215downloads
Model Card

LiteRT models for Android dataset production

Français / English · Charlbi · Documentation 2026-09-21

25 variantes actives, dont 15 avec apprentissage de tête/adaptation. Poids téléchargés séparément de l’APK. Les encodeurs visuels restent figés.

25 current variants, including 15 trainable heads/adapters. Model weights are downloaded separately from the APK. Visual backbones remain frozen.

Résultats illustrés / Visual evidence

Résultats complets FR · Full results EN · Utilisation FR · Usage EN

[image]

5 succès applicatifs, 1 échec RTMDet, 19 variantes à tester. / 5 application passes, 1 RTMDet failure, 19 variants pending. A separate RepViT learning SDK report is documented separately. These are functional checks, not phone speed or accuracy benchmarks.

Current catalogue / Catalogue actif

ConversionHost trainStandalone SDKApp
caformer_s18_learningPASSPENDINGPENDING
convformer_s18N/APENDINGPENDING
convformer_s18_learningPASSPENDINGPENDING
dinov2N/APENDINGPENDING
dinov2_learningPASSPENDINGPENDING
edgenext_small_usi_learningPASSPENDINGPASS
edgenext_x_small_learningPASSPENDINGPASS
edgenext_xx_small_learningPASSPENDINGPASS
efficientformer_l1_learningPASSPENDINGPENDING
efficientvit_samN/APENDINGPENDING
florence2N/APENDINGPENDING
hgnetv2_b0N/APENDINGPENDING
hgnetv2_b0_learningPASSPENDINGPENDING
repvit_m1N/APENDINGPASS
repvit_m1_learningPASSReported PASSPENDING
rfdetrN/APENDINGPENDING
rfdetr_learningPASSPENDINGPENDING
rtmdet_tinyN/APENDINGFAIL
rtmdet_tiny_learningPASSPENDINGPASS
table_transformer_detection_learningPASSPENDINGPENDING
table_transformer_structure_learningPASSPENDINGPENDING
tinyclipN/APENDINGPENDING
tinyclip_learningPASSPENDINGPENDING
vitposeN/APENDINGPENDING
vitpose_learningPASSPENDINGPENDING

Reading the evidence / Lire les preuves

[image]

The reports distinguish host conversion, standalone Android SDK and Vision Dataset Studio integration. Historical converter flags are retained in the original JSON; dated model-card updates link to later app receipts. No physical ARM phone, task accuracy or generalization is claimed.

Les rapports distinguent conversion hôte, SDK Android autonome et intégration à Vision Dataset Studio. Les indicateurs historiques des JSON restent inchangés ; les cartes datées renvoient aux reçus applicatifs ultérieurs. Aucune précision métier, généralisation ni qualification sur téléphone ARM n’est annoncée.

Files and rights / Fichiers et droits

Each models/ folder preserves weights, hash manifests, runtime contracts, conversion/training reports and upstream notices where supplied. Historical fixed-shape versions are in archive. Use a pinned revision and verify hashes before loading. Each upstream licence applies independently; this collection does not grant a blanket licence to third-party weights.

Chaque dossier conserve poids, empreintes, contrats, rapports et attributions disponibles. Utiliser une révision épinglée, vérifier les hashes et consulter les conditions propres au modèle ; aucune licence globale des poids tiers n’est accordée ici.

Application: Vision Dataset Studio · Reproducible figure data.