I’m building a solution that requires providing LLM with some domain-specific knowledge in order for it to answer my questions sufficiently accurate. This domain-specific knowledge is static, it’s like a set of facts to take into account while responding to my prompt.
At the moment, I add this domain-specific information to my prompt as in-context learning, but it has grown quite significantly, so my single question takes more than 6k input tokens.
Does it make sense to use fine-tuning to provide my model with this domain-specific knowledge?
Thanks for a quick reply. Could you give me a little more of explanation behind it?
The way I understand how fine-tuning works is that a developer is supposed to prepare a huge number (e.g., 1000) of prompt-completion pairs that are somewhat similar for a given problem. How should it look in the case of providing the model with domain-specific knowledge? My knowledge is just a long text narrative. There are no prompts and completions within it.