Meta Llama

About

Meta Llama 2 is the next generation of Meta open source large language model. Llama 2 is available for free for research and commercial use and is optimized for assistant-like chat use cases. It is available in different sizes, ranging from 7B to 70B parameters, and includes model weights and starting code for pretrained and fine-tuned language models. Llama 2 is optimized to run locally on Windows and is available through various cloud service providers such as Microsoft Azure, Amazon Web Services, and Hugging Face. It was trained on 2 trillion tokens and has double the context length of its predecessor, Llama 1. Llama 2 has been used by businesses as core ingredients for new generative AI-powered experiences and has received support from various cloud providers and research institutions. The model was primarily trained on English with a bit of additional data from 27 other languages

Features

• Context Length: Llama 2 has a context length of 32,000 tokens, allowing it to consider a significantly larger amount of information when processing and generating text
• Open Source: Llama 2 is open source and available for free for research and commercial use
• Many sizes: Llama 2 is available in 7B, 13B and 70B sizes. The larger the sizes the more powerful the model is.

Limitations

• Safety Measures: Llama 2 introduce safety measures causing it to be less restrictive but still more flexible than OpenAI GPTs.
• Factualness: Factual information should be taken with a grain of salt.
• Mathematical Reasoning: Llama 2 models are not good at mathematical reasoning.
• Reasoning: Llama 2 models still falls behind GPT-4 in reasoning but close to GPT-3.5 performance.

Usage Tips

• System Prompt: Use the system prompt to modify the Mistral behavior. i.e. You are a helpful assistant that responds in concise and clear language.
• Temperature: Increase the Temperature if you want more creativity in the generated text.
• Top P: Reduce the Top P value if you want more control over the generated text. i.e. Top P = 0.1 will only use the top 10% of the most likely words.
• Top K: Reduce the Top K value if you want more control over the generated text. i.e. Top K = 40 will only use the top 40 most likely words.