Basic Information

Model Name: Llama 3 70B Instruct

Developer/Creator: Meta

Release Date: April 18, 2024

Version: 3

Model Type: Large Language Model (LLM)

Description

Overview:

Llama 3 70B Instruct is a state-of-the-art large language model designed for assistant-like chat and natural language generation tasks. It represents a significant leap over its predecessor, Llama 2, offering improved capabilities in reasoning, code generation, and instruction following.

Key Features:

• Advanced reasoning and code generation capabilities
• Improved instruction following and alignment
• Reduced false refusal rates
• Increased diversity in model responses
• Enhanced steerable outputs

Intended Use:

The model is primarily intended for commercial and research use in English, focusing on assistant-like chat applications and various natural language generation tasks.

Language Support:

While primarily designed for English, developers may fine-tune the model for other languages, provided they comply with the Llama 3 Community License and Acceptable Use Policy.

Technical Details

Architecture:

• Decoder-only transformer architecture
• 70 billion parameters
• Grouped Query Attention (GQA) for improved inference efficiency
• 128K token vocabulary for more efficient language encoding
• 8,192 token context window

Training Data:

• Trained on up to 15 trillion tokens
• Continued improvement observed even after training on two orders of magnitude more data than the Chinchilla-optimal amount
• Diverse dataset to enhance model performance and reduce biases

Performance Metrics:

• MMLU score: 0.82
• Quality Index across evaluations: 62
• Output speed: 54.3 tokens per second
• Time to First Token (TTFT): 0.44 seconds

Comparison to Other Models:

• Establishes new state-of-the-art for LLM models at the 70B parameter scale
• Outperforms competing models of comparable size in real-world scenarios, as evaluated by human annotators

Usage

Code Samples:

Ethical Guidelines

Llama 3 70B Instruct adheres to strict ethical considerations:

• Promotes openness, inclusivity, and helpfulness
• Designed to serve a wide range of use cases and backgrounds
• Respects user dignity and autonomy
• Avoids unnecessary judgment or normativity
• Follows a Responsible Use Guide to mitigate potential misuse and critical risks

Licensing

License Type: Llama 3 Community License

• Allows for commercial and research use
• Prohibits use in ways that violate applicable laws or regulations
• Full license details available at https://llama.meta.com/llama3/use-policy/]