cognitivecomputations

/

dolphin-2.9.1-llama-3-8b

like 36

When using the system message as pre-prompt it exposes it in the response even by strictly instructing not to do and the response does not reflect the user instruction at all (or only a small portion of it).
But when I set the system message as assistant role it works perfectly.

I checked against other models and other chat clients to ensure the issue was not there.

I use Ollama as a client on Linux Ubuntu 22.04 with the Q4_0 model.