据Hacker News报道，曾在Meta任职的前研究员Stephen Roller在证词中承认，他通过BitTorrent下载了包含近20万本图书的“Books3”数据集，该数据源自私人种子站点Bibliotik 1。Roller表示，他在2017年至2023年供职于Meta期间下载并分发了这一数据集，随后称因“所有诉讼的压力”而删除了本地副本 1。这一行为被版权方作为关键证据，旨在挑战Meta关于合理使用AI训练数据的辩护立场 1。
尽管法院去年已裁定使用受版权保护的书籍训练Llama模型属于合理使用，但数据获取手段的合法性已成为新的争议焦点 1。法官本月下令要求Meta公开批准从影子图书馆下载数据的员工名单，以厘清数据来源的合规性 1。针对Roller的行为，Meta回应称其未明确说明下载是为了服务于Meta还是其他雇主，试图将此举定性为个人而非公司行为 1。此外，原告方面正寻求披露Meta内部名为“ML Books”的Google Drive文件夹内容及公司读书会记录，以进一步证明存在非AI用途的盗版内容共享情况 1。目前，Meta声称仍在搜索备份服务器日志，且无证据表明相关命令历史文件已被删除 1。涉及Entrepreneur Media的案件计划于2027年春季开庭审理 1。
Stephen Roller, who worked at Meta from 2017 to 2023, admitted in deposition testimony that he downloaded and distributed the Books3 dataset via BitTorrent 1. Roller stated that he deleted his copies "due to all this litigation," a claim cited by rightsholders including Sarah Silverman, Ta-Nehisi Coates, and Entrepreneur Media to challenge Meta's fair use defense regarding AI training data 1. The Books3 dataset comprises nearly 200,000 books sourced from the private torrent site Bibliotik 1. Although a court ruled last year that using these books to train Llama models constituted fair use, the focus of legal scrutiny has shifted to the legality of how the data was originally acquired 1.
The dispute has prompted broader demands for transparency within Meta. In California federal court, plaintiffs plan to question Meta about the number of Books3 copies obtained and the methods used to acquire them 1. Additionally, plaintiffs have requested disclosure of an internal Google Drive folder named "ML Books" and records from company book clubs, alleging evidence of pirated content sharing for non-AI purposes 1. A judge ordered Meta earlier this month to reveal the names of employees approved to download data from shadow libraries 1.
Meta’s response highlights ambiguities in Roller’s admission and ongoing discovery efforts. The company noted that Roller did not specify whether his downloads were intended for Meta or another employer 1. Regarding digital evidence, Meta claims it is still searching backup server logs and asserts there is no proof that command history files were deleted 1. The Entrepreneur Media case related to these issues is scheduled for trial in spring 2027 1.
评论
还没有评论，欢迎留下第一条。