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.

license: apache-2.0 language: