If you think this is even close to gpt3.5turbo quality, try passing code output from OpenAssistant straight to 3.5turbo and see what happens.