Huge Announcement from Meta: Introducing Llama 3.1!

Meta has unveiled the Llama 3.1 family of multilingual large language models (LLMs). Here’s what you need to know:

• The Meta Llama 3.1 family includes pre-trained and instruction-tuned generative models in 8B, 70B, and 405B sizes (text in/text out).
• All models support long context length (128k) and are optimized for inference with grouped query attention (GQA).
• Optimized for multilingual dialogue use cases, outperforming many existing open-source chat models on industry benchmarks.
• Llama 3.1 uses an optimized transformer architecture with SFT and RLHF for alignment, maintaining the same dense structure as Llama 3 for text input and output.
• The Llama 3.1 Instruct Model (Text) is fine-tuned for generating tool calls for search, image generation, code execution, mathematical reasoning, and supports zero-shot tool use.

System Safety:

• Features models and inference guardrails with Llama-Guard-3 8B for multilingual text input.
• Includes prompt guard for direct and indirect prompt injection filtering.
• CyberSec Eval 3 assesses LLM security risks, including prompt injection and code interpreter abuse.
• Introduces the False Refusal Rate (FRR) to measure safety-utility tradeoffs.
• Provides open-source code for further evaluations.