CL-Tuning: A Curriculum Learning-Based Progressive Fine-Tuning Framework for Large Language Models in Social Manufacturing

Fine-tuning large language models (LLMs) can significantly enhance their ability to understand and reason about complex knowledge in the field of social manufacturing, thereby advancing the development of intelligent manufacturing. However, social manufacturing involves a large amount of highly specialized and logically complex knowledge, and current fine-tuning techniques often overlook the hierarchical relationships within this knowledge, leading to suboptimal learning outcomes for LLM. To address this issue, we propose a Curriculum Learning-based Fine-Tuning framework (CL-Tuning), which aims to guide LLMs in progressively learning the complex knowledge of social manufacturing. Specifically, we design a self-paced learned knowledge difficulty measurer to automatically categorize the knowledge in the social manufacturing corpus into multiple difficulty levels. Subsequently, we fine-tune the general LLMs in stages according to the categorized difficulty levels and introduce a knowledge distillation loss to prevent the model from forgetting foundational knowledge from earlier stages during progressive fine-tuning. Comparative experiments and a case study on open-domain QA datasets and a custom social manufacturing QA dataset validate the effectiveness and superiority of the proposed framework in enhancing the model's understanding and reasoning capabilities for complex manufacturing knowledge, demonstrating significant improvements over existing fine-tuning techniques.