A L ightweight LLM- B ased M ethod for G enerating G ame NPC D ialogues, F ocusing on R ole C onsistency and T ask E xecution

The big language model can improve the degree of freedom of the game NPC dialogue, but the lightweight model may still have problems such as role tone distortion, knowledge crossover, and task information omission in a specific worldview. This paper is based on Qwen3-1.7 B-Instruct for 6 GB local devices, using QLoRA to fine-tune the domain, and using dynamic constraint prompts to coordinate the six dimensions of role, world state, scene, task, historical dialogue, and player input. The text of 875 empty knights was cleaned and structured, and 4375 samples were obtained after enhancement, including 440 test sets. The evaluation system includes BLEU, ROUGE, Distinct, generative behavior, and LLM-as-a-Judge. The BLEU and ROUGE of the candidate model lr-5e5-epoch2 are higher than the benchmark model current-reference, but the average output is longer, and the EOS termination rate and Distinct are significantly reduced. DeepSeek Judge completed the analysis of 440 pairs of responses and preferred the benchmark model in the non-parlement sample with a winning rate of 55.71 % to 44.29 %. The results show that the improvement of reference similarity does not necessarily bring about the simultaneous improvement of role consistency and generation stability, and multi-dimensional evaluation helps to make a choice between task fit and role robustness.