Text-to-SQL data synthesis must expand semantic coverage without losing the relational structure of observed queries. Existing SQL offers evidence of plausible join and nesting patterns, but reusing queries alone limits the range of training examples; unconstrained generation, in contrast, risks implausible structures and question–SQL mismatches. We introduce Structure-Constrained Data Synthesis (SCDS), which treats observed query structure as a prior and uses large language models (LLMs) to vary clause-level semantic content. SCDS groups source queries by coarse structural signatures and constructs semi-frozen SQL templates that retain the source FROM/JOIN context while abstracting selected clauses into semantic slots. It grounds intent generation in schema semantics and available values and represents the resulting objectives in a semi-structured intermediate representation. It then instantiates SQL, evaluates its executability diagnostically, and generates the corresponding questions and reasoning descriptions. On BIRD-dev, fine-tuning with the full set of 23,374 synthesized examples achieves a mean execution accuracy of 57.69% across three training seeds. On MaritimeSQLBench, fine-tuning with synthesized examples improves execution accuracy from 47.69% to 56.75% and reduces the execution error rate from 13.33% to 10.77%. The maritime result is measured with the gold schema elements required by each reference SQL. We publicly release MaritimeSQLBench to support further research on maritime Text-to-SQL. These results support structural priors as a useful constraint on Text-to-SQL synthesis. At a matched training size of 9428 examples, SCDS achieves performance comparable to BIRD-train; together with the maritime results, this highlights its potential for domain-specific Text-to-SQL data synthesis.

Maritime logistics relies on documents that combine dense multilingual text, irregular layouts, and task-specific output schemas, requiring operational systems to recover text, structure, location, and semantics. We trained Hi-Dolphin-OCRusing a task-adaptive reinforcement-learning framework that formulates text spotting, structured localization, document parsing, information extraction, and visual question answering as a unified instruction-following problem. The framework combines task-routed verifiable rewards with bidirectional spatial coverage for spotting, a hard degeneration gate for repetitive, symbol-dominated, and overlong responses, and difficulty-aware sampling based on the supervised model’s error distribution. To our knowledge, we also release the first publicly available maritime-oriented multi-task OCR dataset, Maritime-OCR-Bench, comprising 1888 maritime and complex-document examples across five evaluation tasks. On this benchmark, the failure-aware aggregate score increases from 0.7589 for the supervised checkpoint to 0.7900 after reinforcement learning, with the largest differences observed for coordinate-bearing structured outputs. These results support task-aligned reinforcement learning as a practical strategy for unified OCR, while the released dataset provides a reproducible basis for evaluating spatial, structural, and semantic document understanding in maritime-oriented settings.