This performs considerably better than simply running llama.cpp. I've tested with 7b, 13b and 30b parameters alpaca models locally and I couldn't get them to produce accurate instructions to cut a video with ffmpeg, but this one does.

Wonder if it's the 4bit quantization.