Qwen2.5 Coder 7B Instruct hits XQuery-to-SQL limits

This looks less like a model-selection problem and more like a systems problem. Fine-tuning Qwen2.5-Coder 7B on a tiny dataset may improve style consistency a bit, but it will not reliably teach coverage for the combinatorial space of XQuery variants.

• –110 to 120 samples is enough for a narrow formatter, not for robust semantic translation across many XQuery shapes
• –Regex parsing is the wrong foundation here; use a real XQuery parser or intermediate AST/IR, then let the LLM map that structure to SQL
• –Constrained decoding or schema-guided generation will likely reduce missing columns and conditions more effectively than a small LoRA alone
• –Synthetic data generation is probably the highest-leverage move: generate many paraphrased XQuery variants and corresponding SQL under controlled templates
• –A local model like Qwen2.5-Coder 7B is a reasonable base, but the bottleneck is coverage and verification, not raw model capability