SOAEs-DataSuite: Tailored Pre-training Corpus, Fine-tuning Dataset and Benchmark for Large Language Models in the State-owned Assets and Enterprises Domain

Although Large Language Models (LLMs) make revolutionary progress in diverse fields, their direct application in the domain of state-owned assets and enterprises (SOAEs) of China still faces unique challenges due to distinct business contexts, regulatory frameworks, and customization demands. To bridge this gap and address the critical issues of data scarcity, quality, and diversity within SOAEs, we propose SOAEs-DataSuite: a meticulously curated pre-training corpus, finetuning dataset, and comprehensive benchmark tailored for LLMs in the SOAEs domain. This suite encompasses a diverse range of data spanning knowledge question answering, issue-suggestion extraction, and content classification tasks, reflecting the specific demands of this domain. Specifically, our main contributions lie in: (1) the development of a tailored DataSuite, laying the groundwork for specialized SOAE-centric LLM development; (2) the introduction of a model evaluation framework tailored to SOAEs’ unique requirements, leveraging metrics such as accuracy for classification and BLEU/Rouge scores for question-answering and issue-suggestion extraction tasks; and (3) demonstrating the effectiveness of specialized LLMs trained on SOAEs-DataSuite through rigorous experimental evaluations. The successful creation of SOAEs-DataSuite underscores the importance of customized datasets and LLMs in empowering state-owned enterprises with efficient AI solutions, in line with the government’s push for industrial transformation and innovation.