Evaluating Large Language Model Performance in Generating Culturally Authentic Indian Travel Content: A Benchmark Study Across Six Models and Three Prompt Strategies

This study presents a systematic benchmark of large language models for Indian travel content generation. A human-authored benchmark corpus covering 20 Indian destinations across spiritual, heritage, adventure, and nature categories was constructed and used to evaluate six large language models: GPT-OSS-120B, Qwen3-32B, LLaMA3-70B, LLaMA4-Scout, Qwen3.6-27B, and LLaMA3-8B. Each model was evaluated under three prompting strategies: zero-shot, context-enriched, and persona-based, resulting in 360 generated texts. Generated content was evaluated using ROUGE-1, ROUGE-2, and ROUGE-L, Flesch readability measures, keyword coverage, TF-IDF cosine similarity, and word count. Statistical analyses were conducted to examine the effects of model and prompt strategy. In addition, seven evaluators assessed generated texts using a five-point Likert scale across factual accuracy, cultural authenticity, emotional tone, and practical usefulness. The results show that GPT-OSS-120B achieved the highest overall automated composite score and human-evaluation mean, while Qwen3-32B performed strongly in ROUGE-1, readability, and emotional tone. Persona-based prompting achieved the strongest ROUGE and keyword-coverage results, while zero-shot prompting achieved slightly higher TF-IDF cosine similarity. The study also identifies differences between Qwen model versions, demonstrating the importance of task-specific empirical evaluation rather than assuming that newer model versions always perform better.