OpenAI Playground Start & Restart Prompt Injection
Recently I wrote on how OpenAI is introducing structure into the process of prompt engineering and LLM Interaction.
The introduction of structure is being performed by OpenAI on a few fronts:
1️⃣ The use of ChatML to distinguish between human and AI text. OpenAI sees the introduction of ChatML as an opportunity to mitigate and in future solve for malicious prompt injections, as the model will be able to distinguish between instructions from users, developers and own input.
2️⃣ Different modes were introduced, a seen in the graphic below. Each of the modes have a specific endpoint, model and prompt template. Modes can be seen as use-cases or goal-orientated configurations.
3️⃣ A few days ago the OpenAI Playground introduced two new configuration options in the Complete mode of their playground. The option allows for the injection of start and restart text.
Inject Start Text and Inject Restart Text parameters enable:
- The insertion of text at the beginning or
- at the end of the completion.
The example in the image above, shows a chatbot scenario, and Inject Start Text is triggered when the user clicks “Submit”.
A new line with the keyword “AI:” is created and added to the prompt and the completion text mode initiates the generation request.
Inject Restart Text is used to continue the conversational pattern and dialog turns within the prompt.
Hence when completion ends, a new line with the keyword “Human:”, is appended to the prompt.
These parameters are not mandatory, but is useful for chat specific completion tasks.
It also serves an educational purpose in guiding the user to create well formed prompts; especially for a conversational UI use-case.
Lastly, in my view, these developments all form part of OpenAI’s ongoing process of introducing features, structure and user guidance to the playground. And hence the playground is slowly morphing into a LLM development UI.
⭐️ Please follow me on LinkedIn for updates on Conversational AI ⭐️
I’m currently the Chief Evangelist @ HumanFirst. I explore and write about all things at the intersection of AI and language; ranging from LLMs, Chatbots, Voicebots, Development Frameworks, Data-Centric latent spaces and more.
