The rise of code-generating LLMs such as DeepSeek, Qwen, and Meta-LLaMA has improved developer productivity but also increased risks of plagiarism, copyright misuse, and insecure machine-generated code. While AI-text detection is well studied, machine-generated source-code detection especially across multiple languages, LLM families, and OOD conditions-remains underexplored. SemEval-2026 Task 13 addresses this via two subtasks: (A) binary human–machine code detection and (B) multi-class authorship attribution across ten LLM families. For Subtask A, we fine-tune RoBERTa, CodeBERT, GraphCodeBERT, and StarCoderBase-1B, introducing a stratified sampling strategy with class-weighted loss to mitigate imbalance and OOD shifts. For Subtask B, we mitigate the extreme human-class imbalance using undersampling, inverse-frequency weights, syntactic noising, and curriculum-based dual-path training with TinyStarCoderPy and CodeBERT. Both results show that long-context modeling, distribution-aware sampling, and noise-robust training are crucial for reliable in real-world settings. Overall, long-context modeling, distribution-aligned sampling, and lightweight noise-robust training emerge as key factors for reliable machine-generated code detection and authorship attribution.