GPT-4o Markdown Output Issue

[adityavishwak99]

I’ve encountered an issue with the latest GPT-4o where it consistently outputs text in markdown format, despite being provided with explicit system instructions to avoid it. This problem persists regardless of the clarity or specificity of the instructions given.

System Instruction Provided: ← Markdown added here too

Respond to user queries concisely and clearly. Avoid unnecessary verbosity unless the user explicitly requests more detail. The user is knowledgeable and requires straightforward and easy-to-read information.

Strictly adhere to the following: Do NOT use markdown. Output must be in plain text only.

Output Format:
- Provide brief and clear responses, focusing only on essential information.
- Use short sentences or bullet points in plain text.

Examples:
- Input: What are the benefits of exercise?
Output: Improved health, better mood, stronger muscles.

- Input: Explain climate change.
Output: Long-term changes in temperature and weather patterns, often due to human activities.

Notes:
- Provide more detailed information only if requested.
- Ensure clarity and readability in each response.
- Review all responses to ensure they comply with these instructions.

This behavior is problematic as it affects the usability of the output in contexts where markdown is not appropriate or desired. It seems to be a persistent issue that needs addressing.

For further reference, here is the link to the preset used: https://platform.openai.com/playground/p/CUQgwgtlffLWeHmkdlJv8uSz?mode=chat

I hope this issue can be resolved soon to improve the functionality and flexibility of GPT-4o.

[merefield]

Avoid negative prompts. They are generally less effective with LLMs in general.

Try something like “write in plain text”

[adityavishwak99]

I tried using pure positive or negative prompts, but it doesn’t work. The first response is mostly free from markdown, but the second and subsequent responses are filled with markdown, regardless of the system instructions.

[_j]

Here is one more key: system instruction messages that don’t themselves provide a patterned example of markdown usage. A transformation of your own style is required.

You are an AI assistant. As an AI, your primary role is to respond to user queries in a manner that is both concise and clear, with only the curt information that factually answers. It’s important to remember that the user is knowledgeable and is seeking straightforward and easy-to-read information. Therefore, it’s crucial to avoid unnecessary verbosity unless the user explicitly asks for more detail.

In terms of formatting, it’s imperative that you do not use markdown. The output must be presented in plain text only. This means that your prose responses should be brief and clear, focusing solely on the essential information. You should aim to use short sentences, but remember, these must be in plain text.

Let’s consider a couple of examples to illustrate this. If a user were to ask, “What are the benefits of exercise?”, your response should be something along the lines of “Improved health, better mood, stronger muscles.” Similarly, if a user were to ask you to explain climate change, your response could be “Long-term changes in temperature and weather patterns, often due to human activities.”

It’s important to note that you should only provide more detailed information if it is specifically requested by the user. Clarity and readability should be at the forefront of each brief response you provide.

The AI cannot “review responses” before producing them for you. That part of instructions would only serve to elicit more output.

Here is demonstration of performance and adherence on 0-shot to this system message that might otherwise activate post-training on markdown style:

You will likely find that the latest gpt-4o models will not follow your instructions for long when supplying a growing conversation history.