llm-training-data

论文提出 Edit Discovery Optimization (EDO) 问题框架，用于在有限编辑器成本预算下选择性地将人机交互数据交给更强的编辑器（人类专家或更强模型）检查与修正，并提出在线算法 EDIT 在四个推理与事实问答基准上取得更优的成本-错误权衡。

This paper assesses the prevalence of extremist speech in the training data of large language models, focusing on the Dolma corpus, and discusses implications for data curation and model safety.

An update on the escalating problem of AI scraper bots overwhelming websites, discussing residential proxy networks and their impact on the open web.

A personal blog post explaining an anti-crawler browser blocking mechanism, including special notes for Inoreader, Feedly, Vivaldi, and archive.* users.