Tuning to produce higher quality responses

[Anton-V-K]

Has anyone succeeded in tuning the engine so it produces higher quality responses?

Here is a case with the prompt

Is 450 90% of 500?

privateGPT produces something like this:

No, 450 is not 90% of 500. To find 90% of 500, you would multiply 500 by 0.9 (which represents 90% as a decimal). The result would be 450, but this means that 450 is equal to 90% of itself, not 500.

I find such responses somewhat gibberish :)

For instance, GPT-4 is able to correct itself while answering same question:

or even give the correct answer immediately:

Another case with Claude 2.1 (well, this engine is built on another "technology"):

Probably the engine needs extra tuning?
I've started with defaults in settings.yaml:

local:
  prompt_style: "llama2"
  llm_hf_repo_id: TheBloke/Mistral-7B-Instruct-v0.1-GGUF
  llm_hf_model_file: mistral-7b-instruct-v0.1.Q4_K_M.gguf
  embedding_hf_model_name: BAAI/bge-small-en-v1.5