soup

Fine-tuning tool that configures LLM post-training from a single YAML file and command. Using layer streaming, it enables developers to train 8B models on 4GB laptop GPUs with limited memory.