unicornwhodev/Lite_rt_prepared_for_android_dataset_builder

LiteRT Models for Android Dataset Production

A collection of 25 current LiteRT conversion variants, including 15 trainable heads or output adapters, for experimental Android dataset and annotation workflows. Weights are downloaded separately from the application package. Visual backbones remain frozen; the trainable scope is defined by each variant's runtime contract.

Maintained by Unicorn Who Dev. English usage guide · Detailed results.

Recorded qualification

The retained application campaign records 5 PASS, 1 FAIL and 19 PENDING variants. All 15 learning variants passed synthetic host checks. RepViT M1 learning has a separate, report-level Android SDK result; it is not an application pass.

Conversion	Host train	Standalone SDK	App
caformer_s18_learning	PASS	PENDING	PENDING
convformer_s18	N/A	PENDING	PENDING
convformer_s18_learning	PASS	PENDING	PENDING
dinov2	N/A	PENDING	PENDING
dinov2_learning	PASS	PENDING	PENDING
edgenext_small_usi_learning	PASS	PENDING	PASS
edgenext_x_small_learning	PASS	PENDING	PASS
edgenext_xx_small_learning	PASS	PENDING	PASS
efficientformer_l1_learning	PASS	PENDING	PENDING
efficientvit_sam	N/A	PENDING	PENDING
florence2	N/A	PENDING	PENDING
hgnetv2_b0	N/A	PENDING	PENDING
hgnetv2_b0_learning	PASS	PENDING	PENDING
repvit_m1	N/A	PENDING	PASS
repvit_m1_learning	PASS	Reported PASS	PENDING
rfdetr	N/A	PENDING	PENDING
rfdetr_learning	PASS	PENDING	PENDING
rtmdet_tiny	N/A	PENDING	FAIL
rtmdet_tiny_learning	PASS	PENDING	PASS
table_transformer_detection_learning	PASS	PENDING	PENDING
table_transformer_structure_learning	PASS	PENDING	PENDING
tinyclip	N/A	PENDING	PENDING
tinyclip_learning	PASS	PENDING	PENDING
vitpose	N/A	PENDING	PENDING
vitpose_learning	PASS	PENDING	PENDING

PASS applies only to the indicated check. PENDING means unexecuted or unfinished; N/A means no training signature is exposed. The inference-only RTMDet Tiny failure is a feature-map/stride mismatch. Its learning variant is a different graph and passed its recorded application check.

What the checks establish

The application checks used an AOSP API 28 x86_64 software emulator, LiteRT 1.4.2, Select TF Ops 2.16.1 and a synthetic image. They establish execution and, where tested, optimizer updates, checkpoint restoration and resume. Whole-test durations include repeated inference and checkpoint operations and are not inference latency or FPS. No successful case produced a proposal on that fixture.

Physical ARM device qualification, representative task accuracy, generalization, forgetting, memory, thermal behavior and repeated latency measurements remain open. These historical receipts do not qualify a later application binary.

Successful application receipts · Failure receipt · Figure source data.

Download and integrate

Read the selected models/ folder's README, runtime contract and notices first. Download a pinned revision, verify its artifact hashes, and follow its exact preprocessing, tensor layout, labels and coordinate transform. Keep the original weights and save learned checkpoints separately. Evaluate a candidate on authorized, representative data before activation.

from huggingface_hub import hf_hub_download

contract = hf_hub_download(
    repo_id="unicornwhodev/Lite_rt_prepared_for_android_dataset_builder",
    revision="36026262693de56b2cf45a6337a405297bfcfff6",
    filename="models/edgenext_xx_small_learning/runtime_contract.json",
)

This example retrieves a contract; inference and learning use the model-specific instructions. Earlier fixed-shape variants are retained in the historical archive and are excluded from the 25 current variants.

Rights and intended use

Each upstream model and component retains its own licence and notices. This collection grants no blanket licence to third-party weights. Preserve provenance and respect each model's permitted uses before download, adaptation or redistribution.

Application reference: Vision Dataset Studio.

Documentation and revision

Card updated in English on 5 October 2026, from the public repository at 36026262693de56b2cf45a6337a405297bfcfff6 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.