CodyDroid Model Performance

The CodyDroid Evaluation Dataset is curated to benchmark the Android SDK code generation capabilities of two lightweight language models: CodyDroid Model-1 (deepseek-coder-5.9M-kexer) and CodyDroid Model-2 (codegen-175K-mono-java). Each entry in the dataset consists of a natural language description of an Android development task, paired with code outputs from both models, a reference solution crafted by a developer, and a detailed qualitative analysis. The tasks span typical Android programming challenges involving SDK usage, Jetpack libraries, UI handling, and system services. The dataset is designed to support comparative evaluation, model fine-tuning, and error analysis in the context of mobile app development, offering a unique resource for studying LLM performance in constrained-code and domain-specific scenarios.