A proposed class-action lawsuit filed by Oregon author Elizabeth Lyon accuses Adobe of training its SlimLM AI model on pirated books, including her guidebooks, through the SlimPajama-627B dataset derived from the RedPajama collection containing Books3.
Adobe has pursued extensive development in artificial intelligence over recent years. The company launched multiple AI services starting in 2023, with Firefly serving as its AI-powered media-generation suite designed for creating images, videos, and other media content from text prompts and inputs.
SlimLM represents a series of small language models that Adobe has optimized specifically for document assistance tasks on mobile devices. These models enable functions such as summarizing documents, extracting key information, and providing contextual help directly within mobile applications.
Adobe states that it pre-trained SlimLM using the SlimPajama-627B dataset. Cerebras released this dataset in June 2023 as a deduplicated, multi-corpora, open-source resource intended for training large language models. The dataset aggregates various text sources after removing duplicates to improve training efficiency and model performance.
Elizabeth Lyon, who specializes in guidebooks for non-fiction writing, initiated the lawsuit claiming that Adobe incorporated pirated versions of numerous books, including her own works, into the training process for SlimLM. The legal action seeks class-action status to represent other affected authors.
The lawsuit details how the SlimPajama dataset originated from the RedPajama dataset, which includes the Books3 collection comprising 191,000 books. Reuters first reported on the filing. The complaint states verbatim: “The SlimPajama dataset was created by copying and manipulating the RedPajama dataset (including copying Books3).” It continues: “Thus, because it is a derivative copy of the RedPajama dataset, SlimPajama contains the Books3 dataset, including the copyrighted works of Plaintiff and the Class members.” Lyon argues that her copyrighted materials appeared in this pre-training data without her consent or compensation.
Books3 has emerged repeatedly in legal disputes within the AI sector, as developers have utilized it to train generative AI systems. The collection contains digitized texts from various genres and authors, making it a comprehensive but contentious training corpus. RedPajama, which incorporates Books3, has also faced mentions in multiple court cases.
