Background: Clinical interpretation of endoscopic reports is often limited by heterogeneous documentation and shortages of specialist expertise. We aimed to develop a multicentre clinical benchmark and a Chinese large language model (LLM) report-generation framework to improve endoscopic report understanding and structured information extraction. Methods: We built EndoAtlas from de-identified electronic health records from five tertiary hospitals in China (N=6161), including surgical reports, pathology reports, and operative diagnoses. We evaluated in-context learning (ICL; 3–10 shots) and low-rank adaptation (LoRA) fine-tuning for report-style generation and structured field extraction. Performance was measured with precision, recall, F1, and BLEU-4. Findings: Increasing ICL demonstrations improved performance across models; for example, Doubao improved from F1 25·84 (3-shot) to 31·77 (10-shot) and BLEU-4 from 57·03 to 69·00, and DeepSeek improved from F1 25·78 to 32·56 with BLEU-4 up to 66·91. In report-generation evaluations, models achieved robust performance for surgical and pathology report generation (F1 64·00–70·95; BLEU-4 77·68–81·62). GLM4 achieved the highest performance for surgical report generation (F1 67·94; BLEU-4 81·41), while DeepSeek achieved the highest performance for pathology report generation (F1 70·95; BLEU-4 81·62). Interpretation: EndoAtlas provides a multicentre, extensible benchmark for endoscopic report understanding and structured report generation. Report-style generation with targeted LLM adaptation could facilitate more standardised endoscopic interpretation and scalable clinical decision assistance, particularly in resource-limited settings.