Unveiling impact of domain knowledge and data scale on open-source large language model specialization in anaerobic...

Exploring the integration of domain knowledge and its data scale holds critical value in enhancing the understanding of anaerobic digestion (AD) by open-source large language models (LLMs). This study develops an automated agent system to extract high-quality AD question-answer pairs from literature and fine-tunes three open-source LLMs. Expert evaluations reveal the fine-tuned Llama3.1-8B-AD (LAD) demonstrates notable professional competitiveness in specialized AD domains, exhibiting a knowledge comprehension level approaching that of GPT-4 in specific tasks (0.67 vs. 0.68). Notably, LAD demonstrates competitive or superior performance in advanced domains like additives and microbial knowledge. Additionally, fine-tuning with complete training datasets notably improves professional understanding (0.60 to 0.67), particularly in medium-to-high-difficulty questions. In contrast, training with only 50% of the data leads to unstable foundational knowledge hallucinations, underscoring the necessity of comprehensive domain data for deep understanding and reasoning in AD. Overall, this study provides a reference paradigm for the future development of large models for anaerobic digestion in the bioenergy field and offers new perspectives for the construction of intelligent bioenergy systems.