Event
Llama 2 und die Frage nach offener KI
On 18 July 2023 Meta introduced Llama 2, a family of pretrained language models and variants adapted for dialogue. The research paper describes models ranging from 7 billion to 70 billion parameters. Meta made model weights available for use and redistribution under its own license, which also permitted commercial use subject to conditions. Sources
The research paper reports two trillion tokens for pretraining. It also describes a trained 34-billion-parameter model whose release was initially deferred because safety testing was incomplete. A 2024 Stanford assessment found that individual data sources were not sufficiently broken down for its disclosure criterion. Research paper and assessment
What does “open” mean here? Meta called Llama 2 open source. The Open Source Initiative disagreed, citing restrictions on certain users and fields of use. Under specified conditions, very large providers had to request an additional license. Availability of model weights alone also does not mean all training data was published. License and response
Sources
- Ankündigung, Forschungsarbeit, Lizenz und Stellungnahme.
- GPT-4 im selben Jahr: ein gesondertes Modell und eine andere Veröffentlichung.
Offen
Umfang der einzelnen veröffentlichten Trainingsmaterialien und Änderungen an Zugangsbedingungen prüfen. / Check the scope of released training materials and changes in access terms.