Make sure that the Response Length (tokens) are set to max this was tested with Pygmalion-7b-4bit-32g-GPTQ-Safetensors and it gives best resoults with that model, gpt should work too.