CMV-LLM: China's mainstream values large language model
Overview
Existing large language models often exhibit significant value bias, making them misaligned with the mainstream values of their respective nations. To address this issue, we explore how to align a large language model's ideological stance with their national mainstream values. We introduce CMV-LLM, a model that aligns with China's mainstream values, by carefully curating large-scale Chinese datasets and fine-tuning. To validate its effectiveness, we evaluate the model using a national values assessment dataset, comparing it with several popular LLMs.
The results demonstrate that our CMV-LLM generates content more accurately reflecting China's mainstream values, providing a novel approach and practical insights for building large language models aligned with national stances.