Two class motion lawsuits had been filed by authors in opposition to Meta in 2023 claiming misuse of their copyrighted books as coaching materials for Llama. The courts haven’t been terribly sympathetic to the authors. The primary case, Kadrey et al. v. Meta Platforms, was filed in July 2023. The second, Chabon v. Meta Platforms, was filed in September 2023. The 2 circumstances had been mixed, and the mixed case was dismissed on all however one depend, direct copyright violation, in December 2023. It nonetheless dragged on with amended complaints for an additional yr; in September 2024 the choose ordered “Based mostly on earlier filings, the Court docket has been beneath the impression that there’s no actual dispute about whether or not Meta fed copyrighted works to its AI packages with out authorization, and that the one actual authorized query to be offered at abstract judgment is whether or not doing so constituted ‘honest use’ throughout the that means of copyright regulation.” Abstract judgement was scheduled for March 2025, though there was loads of exercise from each side by means of February 2025.
Beneath I’ll talk about the progress Meta AI has made with the Llama household of fashions because the fall of 2023. Be aware that they’re not simply language fashions. Some fashions are multi-modal (textual content and imaginative and prescient inputs, textual content output), and a few can interpret code and name instruments. As well as, some Llama fashions are security parts that determine and mitigate assaults, designed for use as a part of a Llama Stack. The next mannequin write-ups are condensed from Meta AI’s mannequin playing cards.
Llama Guard 1
Llama Guard is a 7B parameter Llama 2-based input-output safeguard mannequin. It may be used for classifying content material in each LLM inputs (immediate classification) and in LLM responses (response classification).