(Optional) Use an Arm AI Portal model without a validated example

Check the project and tools

Use this advanced workflow only when the model doesn’t match one of the validated examples. The model might need different files, tokenization, or prompt formatting. It might also need different tensor mapping, runtime APIs, or output decoding.

Start by returning to the original starter project and verifying the files and tools used by the workflow:

    

        
        

cd "$HOME/text-to-text-android/ai-portal-android-app-text-to-text"
test -f scripts/inspect_android_model.py
test -x gradlew
git rev-parse --show-toplevel
java -version
adb get-state
../.hf-venv/bin/python -c \
    'import huggingface_hub; print(huggingface_hub.__version__)'
  

    

    

        
        

$TEXT_MODEL_WORKSPACE = Join-Path $env:USERPROFILE "text-to-text-android"
Set-Location (Join-Path $TEXT_MODEL_WORKSPACE "ai-portal-android-app-text-to-text")
Test-Path scripts\inspect_android_model.py
Test-Path gradlew.bat
git rev-parse --show-toplevel
java -version
adb get-state
..\.hf-venv\Scripts\python.exe -c `
    "import huggingface_hub; print(huggingface_hub.__version__)"
  

    

The output identifies the project root, Java version, connected device, and huggingface_hub version. Both Windows path checks should return True.

Download and inspect the model package

Run the following commands to download and install the model package:

    

        
        

printf 'Arm model repository ID: '
read -r MODEL_SOURCE
printf 'Runtime: '
read -r MODEL_RUNTIME
printf 'Workload: '
read -r MODEL_WORKLOAD

export MODEL_ID="${MODEL_SOURCE##*/}"
export MODEL_DIR="../model-custom-$MODEL_ID"
../.hf-venv/bin/hf download "$MODEL_SOURCE" --local-dir "$MODEL_DIR"

../.hf-venv/bin/python scripts/inspect_android_model.py \
    --model-id "$MODEL_ID" \
    --model-source "https://huggingface.co/$MODEL_SOURCE" \
    --runtime "$MODEL_RUNTIME" \
    --workload "$MODEL_WORKLOAD" \
    --local-model-dir "$MODEL_DIR"
  

    

    

        
        

$MODEL_SOURCE = Read-Host "Arm model repository ID"
$MODEL_RUNTIME = Read-Host "Runtime"
$MODEL_WORKLOAD = Read-Host "Workload"

$MODEL_ID = Split-Path -Leaf $MODEL_SOURCE
$MODEL_DIR = Join-Path .. "model-custom-$MODEL_ID"
..\.hf-venv\Scripts\hf.exe download $MODEL_SOURCE `
    --local-dir $MODEL_DIR

..\.hf-venv\Scripts\python.exe scripts\inspect_android_model.py `
    --model-id $MODEL_ID `
    --model-source "https://huggingface.co/$MODEL_SOURCE" `
    --runtime $MODEL_RUNTIME `
    --workload $MODEL_WORKLOAD `
    --local-model-dir $MODEL_DIR
  

    

Enter the complete Arm Hugging Face repository ID, runtime value, and workload when prompted. Use executorch, litert-lm, litert, or onnxruntime for the runtime. Use text-generation or text-embedding for the workload.

The script creates android_model_config.json, model-context/model-summary.json, and a directory of copied metadata. Review the summary and confirm that it identifies the intended Android artifact, tokenizer, configuration, and template files.

Give the coding agent local evidence

Open the coding agent in the starter project and ask it to do the following:

1. Read android_model_config.json, model-context/model-summary.json, copied metadata, and the model card.
2. Read inference/RuntimeRunner.kt, the task-specific runner interface, and inference/RuntimeRunnerFactory.kt.
3. Read model_catalog.json and the matching adapter stub.
4. Update one catalog entry, the matching adapter, and the factory registration needed for the selected runtime value.
5. Add a pinned official Android runtime dependency.
6. Report any required local AAR or native library before assuming its location.
7. Validate all required package files under filesDir/models/<model-id>/ and keep all resolved paths inside that directory.
8. Implement tokenization, prompt formatting, and runtime invocation.
9. Implement output decoding, stop-token handling, and timing based on the documented model evidence.
10. Build the debug APK and report the exact model files and application-private directory needed on the phone.

The agent must not embed access tokens or silently select another runtime. It must not invent dependency coordinates or claim success while a required AAR or native library is missing.

Review and validate the result

Review every generated source and build-file change. After review, run the build commands from the application repository:

    

        
        

./gradlew :app:assembleDebug :app:lintDebug
  

    

    

        
        

.\gradlew.bat :app:assembleDebug :app:lintDebug
  

    

Copy the exact files reported by the agent into the catalog’s application-private model directory. After copying the files, test representative inputs against a known reference for the package.

A successful build proves that the source compiles. However, it doesn’t prove that tokenization, tensor mapping, or decoding is correct.

What you’ve accomplished

You’ve inspected another Arm AI Portal package, supplied a coding agent with local model evidence, and identified the build and device checks needed to validate a generated adapter.

You can now extend the workflow to use models without a validated example.