.png){kind=icon}
.png){kind=icon}
Return
Cognition-enhanced instruction framework: Accelerating structured battery knowledge extraction with low-parameter models
.png){kind=icon}
PRE
AI
DOI：10.1016/j.jechem.2025.07.056.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The explosive growth of lithium-ion battery literature has led to severe knowledge overload, challenging researchers’ ability to efficiently extract structured information. While large language models (LLMs) offer considerable potential for automating this task, their practical application in scientific domains is nonetheless constrained by high application programming interface (API) costs and computational resources required for fine-tuning. To address these limitations, a cognition-enhanced instruction framework (CEIF) is proposed, wherein a high-performance teacher model (such as DeepSeek-R1) provides dynamic feedback, prompt refinement, and training data optimization to guide the learning process of low-parameter models. Experimental results demonstrate that the low-parameter models (6B–9B) optimized via the CEIF achieve approximately 85% accuracy in battery literature extraction tasks, approaching the performance of GPT-4 while requiring only a single NVIDIA RTX 3090 GPU. Furthermore, the emergence of an “Aha moment” characterized by rapid performance improvement during specialized learning is observed, offering novel theoretical insights for the design and optimization of domain-specific models.
Keywords:
lithium-ion batteries
large language models
knowledge extraction
instruction tuning
low-parameter models
cognitive enhancement
Journal
IF:
14.9
Papers:
6.2K
Citations:
4.5W