“It is not necessary to have domain-expert labelers for every specific task, such as using FFmpeg, because the base model's generalization from pre-training on documentation and code is sufficient.”