'LLaMA2 Chatbot' allows anyone to try out Meta's large-scale language model 'Llama 2' for free from their browser

Meta has released the ' LLaMA2 Chatbot, ' a large-scale language model (LLM) called '

https://llama2.ai/

The menu on the left side of the screen allows you to change the model and parameters, and delete the chat history.

There are three models to choose from: 'LLaMA2-70B' with 70 billion parameters, 'LLaMA2-13B' with 13 billion parameters, and 'LLaMA2-7B' with 7 billion parameters. The more parameters there are, the better the performance.

Below the model selection area, you can freely change three parameters: ' Temprature ' (the higher the value, the more random the output), ' Top P ' (the higher the value, the more diverse the word selection), and ' Max Sequence Length ' (the maximum length of the sentence). The default parameters are 'Temprature' at '0.01', 'Top P' at '0.90', and 'Max Sequence Length' at '2048'.

You can also delete your chat history by clicking 'Clear Chat History.'

To chat, simply enter text into the text box at the bottom of the screen.

It supports Japanese input, but the accuracy is questionable compared to ChatGPT or Bard. Also, when chatting in Japanese, sometimes the chatter would start responding in English.

Furthermore, as the conversation continues, it looks like this:

The 'LLaMA2 Chatbot' was developed by Rajko Radovanović, who is researching AI and LLM at venture capital firm Andreessen Horowitz .

Radovanović writes about 'LlaMA 2,' 'We have found that performance is comparable to GPT-3.5, especially for creative tasks and interactions, but with a much smaller number of parameters.'

The source code for the LLaMA2 Chatbot is available on GitHub.

GitHub - a16z-infra/llama2-chatbot: LLaMA v2 Chatbot
https://github.com/a16z-infra/llama2-chatbot

Additionally, you can deploy your own LlaMA 2 using the source code and Cog below.

GitHub - a16z-infra/cog-llama-template: LLaMA Cog template
https://github.com/a16z-infra/cog-llama-template

The LlaMA 2 API token created with Replicate, which allows for easy deployment of AI models, is also publicly available.