Mastering Gemini: Innovative Strategies for Hebrew Language Model Optimization
Some Prompt Engineering and Working with Gemini In our work, we often use “Act As” prompts to make the model behave like a certain character. Usually,...
Some Prompt Engineering and Working with Gemini
In our work, we often use “Act As” prompts to make the model behave like a certain character. Usually, we start by trying to find ready-made prompts and then adjust them according to conversations with the bot. Here are some sources for initial prompts:
- You can ask similar bots on Character.AI to tell us their prompt (they don’t always agree).
- You can ask another language model to write a prompt for us, for example, if working with Gemini, you can ask GPT4 to write the prompt.
- When I’m working on complex tasks that are very important to me for the continuation of the work and I want to get good results, I use a “super prompt,” a really long prompt that I write at the beginning of the conversation to make the model work better on my task. Here’s an example from an article on Medium. It’s really long, so you don’t have to read it all, you can just copy it to the beginning of the conversation: gist.githubusercontent.com/codewithbernard/237c572b34f18cc6ed3d7f029192fb8f/raw/0e032163ba1915f0fdc4d63f73413452de2a9153/chatgpt-builder.txt
- Another option is to use repositories of example prompts, for example: https://github.com/f/awesome-chatgpt-prompts
When I start the conversation, there’s a fairly high likelihood that the model will think it needs to start with an example, especially with Gemini, where you currently can’t set a context. The solution I’m currently using is to ignore the model’s first message — when working with the API, I simply don’t show it to the user, and start with my own opening message. After that, the model understands that it’s in a conversation and behaves like the character we defined for it.
Attaching a screenshot for demonstration, the model decides it needs to give an example, which of course I don’t show to the user. The screenshot is from Gemini’s Playground (which does not support RTL).
I’d love to hear what you thought or any better ideas you’re using.
I’d love to hear more insights or questions you have about using language models in Hebrew.
If you want to see more posts about our work at Tov-Tech, feel free to follow us on LinkedIn: https://www.linkedin.com/in/raz-hadas/