Touvron et al. (Meta) • 2023

Abstract

Meta's release of Llama 2, a collection of pretrained and fine-tuned language models ranging from 7B to 70B parameters. The paper details training methodology, including Ghost Attention for multi-turn consistency and extensive safety testing, establishing patterns for open-weights model development.

Warum Es Wichtig Ist

• Broadened access to capable open-weights models
• Detailed RLHF and safety methodology
• Reference architecture for open-source development