Abstract
Industrial decarbonization involves knowledge-intensive tasks that require engineers and practitioners to access, interpret, and integrate heterogeneous information from technical literature, policy documents, standards, and domain reports. General-purpose large language models often struggle in this setting because of limited domain alignment and weak factual grounding. This study presents Mini-CarbonGPT, a resource-efficient domain language model for industrial decarbonization knowledge work. The system combines parameter-efficient fine-tuning using Low-Rank Adaptation (LoRA) with retrieval-augmented generation (RAG) to couple domain adaptation with evidence-grounded response generation under limited computational resources. A multi-source knowledge pipeline was constructed using 50,000 general instructions, 5,382 domain-specific instructions, 6,096 research papers, and approximately 60,000 curated encyclopedia entries. To evaluate performance in engineering-oriented knowledge tasks, we developed a benchmark containing 700 objective questions and 249 open-ended questions across five carbon-neutrality subdomains. Results show that Mini-CarbonGPT achieved an average objective accuracy of 80.57%, outperforming the baseline model and showing robust performance in factual accuracy, terminology coverage, and response completeness. The proposed framework demonstrates how lightweight adaptation and external knowledge retrieval can be integrated to support industrial knowledge services in complex engineering domains. The study provides a transferable approach for building domain language models for industrial applications where knowledge reliability, deployability, and reproducibility are critical.

This publication has 19 references indexed in Scilit: