Model Overview

Undi95/Meta-Llama-3-8B-hf is an 8 billion parameter instruction-tuned model from Meta's Llama 3 family, designed for generative text and code. It leverages an optimized transformer architecture and Grouped-Query Attention (GQA) for efficient inference. The model was trained on over 15 trillion tokens of publicly available data, with fine-tuning incorporating publicly available instruction datasets and over 10 million human-annotated examples. Its knowledge cutoff for pretraining data is March 2023.

Key Capabilities

• Dialogue Optimization: Specifically instruction-tuned for assistant-like chat applications.
• Enhanced Performance: Outperforms many open-source chat models on standard industry benchmarks across general reasoning (MMLU, AGIEval), knowledge reasoning (TriviaQA-Wiki), and reading comprehension (SQuAD, DROP).
• Safety and Refusal Mitigation: Developed with extensive red teaming and adversarial evaluations to optimize helpfulness and safety, significantly reducing false refusals compared to previous Llama models.
• Code Generation: Demonstrates strong performance in code generation benchmarks like HumanEval (62.2% for the instruction-tuned 8B model).

Intended Use Cases

• Commercial and Research: Suitable for a wide range of applications in English.
• Assistant-like Chat: Ideal for building conversational AI agents and chatbots.
• Natural Language Generation: Adaptable for various text generation tasks, including summarization, question answering, and creative writing.
• Code Assistance: Can be used as a coding assistant, with evaluations showing comparable or safer behavior to models of equivalent coding capability.