I don’t know why, but maybe the dev team overtrained the Markdown-formatted data for the Gemini model.
It started adding double spaces, often used in coding, to prose. This is bad for writing because users have to remove the double spaces themselves.
I think the dev team needs to retrain the model to distinguish between different kinds of whitespace characters.
One more thing: don’t train the model to use 4 spaces in code; it consumes a lot of tokens; 2 spaces are enough.
Hi @RandomGuy,
Which model you are using currently or it is happening with all models?? Would you mind providing a example screenshot to understand the issue better.
Thanks
Some examples using Gemini 2.0 Flash.
I asked Gemini to describe some Python code and output it in markdown format. Here are some excerpts showing the issue.
* **Agent Creation:** Creates a `ReActAgent` from the defined tools and a system prompt that instructs the agent on how to use the tools. The system prompt emphasizes the need to use the tools and avoid relying on prior knowledge.
* Returns the created `ReActAgent` and the document summary.
* **Top Agent Creation:** Creates the `ReActAgent` using the `CustomObjectRetriever` as the `tool_retriever`. The system prompt instructs the agent to always use the tools provided.
* It is decorated with `@cached_property` so it's only built once.
Well, it happens for all models; just prompt it to write anything. It often likes to add two spaces after a dot “.”, or in some other places.
Yes, and it has for awhile, its specifically when calling the API’s I had to build several scripts to fix it, I think it has something to do with the UI being used when the devs were working with the model because if you use different UIs it will strip a lot of it and render correctly but if you just are trying to work from the API response its a constant issue
Hey guys, Thanks for flagging. Already escalated to internal team. Hope to see a fix soon.