模型文化知识在训练中大量丢失，需从数据源头改进。
The Culture Funnel: You Can't Align What isn't in the Data
- 构建多维标签框架，发现文化信号在后训练阶段显著衰减
- 地理集中、任务专精的数据主导模型训练，导致文化代表性失衡
- 提升数据多样性可增强模型跨文化表现，适合关注AI公平性的研究者
当前文化对齐方法依赖推理阶段干预，假设模型已具备充分文化知识。我们提出现代大模型训练存在文化数据漏斗问题。通过在预训练、微调、对齐和推理数据集上使用多维标签框架分析，发现显式文化信号在后训练阶段急剧下降，而地理集中、任务专精的数据占据主导。多语言性虽提升文化知识的地理多样性，但无法保证均衡表征。我们的标签体系显著提升下游文化基准性能，表明进步需转向训练数据管道的优化。为促进后续研究，我们发布了包含560万样本的文化标记数据集，地址：https://huggingface.co/datasets/CohereLabs/CultureMarkers。
原文摘要 · Abstract (English)
Current cultural alignment approaches focus on inference-time interventions, assuming models already contain sufficient cultural knowledge. We argue modern LLM pipelines suffer from a cultural data funnel. Using a multidimensional tagging framework across pretraining, fine-tuning, alignment, and reasoning datasets, we show explicit cultural signals decline sharply during post-training, while geographically concentrated, task-specialized data dominates. Multilinguality enhances geographic diversity of cultural knowledge but does not ensure balanced representation. Our tags improve downstream cultural benchmark performance, demonstrating that advances require shifting focus in training data pipelines. To facilitate future research, we release our culturally tagged dataset with 5.6M samples at https://huggingface.co/datasets/CohereLabs/CultureMarkers.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。