A brief story of LLama
The LLaMA base model was released in February 2023 and has since seen a number of new fine-tuned models released. We will compare the LLaMA base model to the Alpaca, Vicuna, Koala-distill, GPT4-x-Alpaca, and WizardLM models.
Model size: The LLaMA base model comes in four sizes: 7B, 13B, 33B, and 65B. The Alpaca, Vicuna, Koala-distill, and GPT4-x-Alpaca models come in two sizes each: 7B and 13B. The WizardLM model is only available in a 7B size.
Training data: The LLaMA base model was trained on various data sets. The Alpaca model was trained on 52k GPT-3 instructions. The Vicuna model was trained on 70k ChatGPT conversations. The Koala-distill model was trained on 117k cleaned ChatGPT conversations. The GPT4-x-Alpaca model was trained on 20k GPT4 instructions. Finally, the WizardLM model was trained on 70k instructions synthesized with ChatGPT/GPT-3. The OpenAssistant LLaMA model was trained on 600k human interactions collected from OpenAssistant Conversations.
Tools: The OpenAssistant team released software that allows users to run LLaMA models locally. This is a great way to get hands-on experience using LLaMA models for natural language processing tasks.