On-Premises Question Answering for Turkish IT Service Management Tickets: Synthetic Data Generation, QLoRA Fine-Tuning, and NPU-Accelerated Inference—A Preliminary Feasibility Study

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Enterprise information technology (IT) service management (ITSM) requires accurate Turkish-language responses while limiting exposure of sensitive support records. This preliminary feasibility study presents a local, on-premises question-answering workflow that integrates synthetic data generation, quantized low-rank adaptation (QLoRA), multi-method evaluation, and neural processing unit (NPU) inference. From 149,674 IT support records, a local qwen2.5-coder:32b teacher model generated 48,876 question–answer pairs passing automated filters; 48,757 were retained after formatting. The v2 corpus contained more numbered steps (88.7% versus 6.7%) and specific commands (62.5% versus 0.5%) than v1; these descriptive comparisons did not control for experimental generation conditions. Qwen2.5-7B and Qwen2.5-3B were fine-tuned using ticket-grouped training and validation splits to prevent source-ticket overlap. On a ten-scenario benchmark under greedy decoding (without sampling), ROUGE-1 rose from 0.2603 to 0.3763 for the 7B model and from 0.2048 to 0.4010 for the 3B model after fine-tuning. On a 30-question evaluation set (10 benchmark scenarios and 20 validation records), mean G-Eval scores under a shared local judge were 3.78 and 3.76, and pairwise Arena comparisons favored 7B in 15 cases and 3B in 10, with five ties. Following conversion to OpenVINO INT4, single post-warm-up measurements on an Intel Core Ultra 7 366H showed NPU-enabled throughput of 19.3 versus 13.4 tokens/s for 3B and 10.6 versus 7.4 tokens/s for 7B compared with CPU-only inference. Answer correctness was not validated by independent IT experts. These results therefore demonstrate a feasible local workflow only; the small evaluation set, shared teacher and judge model family, structural rather than correctness-based rubric, and single-run hardware measurements limit generalizability.