Experimental Qwen2.5-Coder-7B models distilled from gpt-oss-20b with 100 responses. A lite series exploring data-efficient knowledge transfer.