Main takeaway: Google open-sourced TabFM, a zero-shot table model that reads tables and outputs predictions without training or feature engineering. It uses alternating row–column attention and row compression, was pretrained on billions of synthetic tables, outperformed tuned supervised methods on 51 datasets, is on Hugging Face/GitHub, and will be integrated into BigQuery for ai.predict SQL calls.
币界网消息,谷歌开源了表格零样本大模型TabFM,用户只需将包含历史数据的表格与待预测的新行合并输入,TabFM就能直接「读表」并输出预测结果,跳过传统模型训练和特征工程。该模型融合了行列交替注意力与行向量压缩机制,能处理更庞大的数据集。谷歌使用结构因果模型生成数亿个虚拟表格来预训练TabFM。在包含51个真实数据集的权威评测中,模型的综合预测性能超越了经过精细人工调参的传统监督算法。TabFM已在Hugging Face和GitHub开源,并将直接集成到谷歌云数据库BigQuery中,未来用户只需使用简单的SQL查询命令(ai.predict)即可调用AI进行预测。