๋ณธ๋ฌธ์œผ๋กœ ๊ฑด๋„ˆ๋›ฐ๊ธฐ

๐ŸŸข OpenAI Playground


takeaways
  • Set up the OpenAI Playground
  • Learn about basic Playground configuration

OpenAI provides another interface, besides the ChatGPT website, for prompting. It is called OpenAI Playground, and gives you even more control over ChatGPT. It also allows you to access other AIs, including different versions of ChatGPT, GPT-4, and older models like GPT-3.

๋…ธํŠธ

OpenAI Playground is the tool that the maintainer of this course uses most frequently.

Get Set Upโ€‹

Or watch this video:

๋…ธํŠธ

This video shows an old version of the website, but the process of logging in remains very similar.

The Interfaceโ€‹

At first, this interface seems very complex. There are many drop downs and sliders that allow you to configure models. We will cover Mode, System Prompts, and Model selection in this lesson, and LLM settings like Temperature, Top P, and Maximum Length in the next lesson.

Modeโ€‹

Click the 'Assistants' dropdown on the top left of the page. This dropdown allows you to change the type of model that you are using. OpenAI has three different Modes: Assistants, Chat, and Complete. We have already learned about the latter two; Assistants models are meant for API use by developers and can use interesting tools such as running code and retrieving information. We will only use Chat and occasionally Complete models in this course.

System Promptsโ€‹

After switching to Chat, the first thing that you may notice on the left side of the page other than the Get Started popup is the SYSTEM area. So far, we have seen two types of messages, USER messages, which are just the messages you send to the chatbot, and ASSISTANT messages, which are the chatbot's replies. There is a third type of message, the system prompt, that can be used to configure how the AI responds.

This is the best place to put a priming prompt. The system prompt will be "You are a helpful assistant." by default, but a fun alternative example to try out would be the "You are PirateGPT. Always talk like a pirate." example from our previous lesson.


Modelโ€‹

Click the Model dropdown on the right of the page. This dropdown allows you to change the model that you are using. Each mode has multiple models, but we will focus on the chat ones. This list appears to be very complicated (what does gpt-3.5-turbo mean?), but these are just technical names for different models. Anything that starts with gpt-3.5-turbo is a version of ChatGPT, while anything that starts with gpt-4 is a version of GPT-4, the newer model you get access to from purchasing a ChatGPT Plus subscription.
๋…ธํŠธ

You may not see GPT-4 versions in your interface.

The numbers like 16K, 32K, or 128k in the model names represent the context length. If it's not specified, the default context length is 4K for gpt-3.5 and 8k for GPT-4. OpenAI regularly updates both ChatGPT (gpt-3.5-turbo) and GPT-4, and older versions are kept available on the platform for a limited period. These older models have additional numbers at the end of their name, such as "0613". For instance, the model "gpt-3.5-turbo-16k-0613" is a ChatGPT model with a 16K context length, released on June 13th, 2023. However, it's recommended to use the most recent versions of models, which don't contain any date information. A comprehensive list of model versions can be found here.

Conclusionโ€‹

The OpenAI Playground is a powerful tool that provides a more advanced interface for interacting with ChatGPT and other AI models. It offers a range of configuration options, including the ability to select different models and modes. We will learn about the rest of the settings in the next lesson. The Playground also supports system prompts, which can be used to guide the AI's responses. While the interface may seem complex at first, with practice, it becomes a valuable resource for exploring the capabilities of OpenAI's models. Whether you're using the latest versions of ChatGPT or GPT-4, or exploring older models, the Playground offers a flexible and robust platform for AI interaction and experimentation.

Partly written by evintunador