MakazhanAlpamys/Soup

Soup is a command-line tool for fine-tuning large language models from a single YAML configuration file.

The tool addresses the complexity of LLM fine-tuning by eliminating configuration overhead and enabling training on consumer-grade hardware. It supports multiple training approaches including supervised fine-tuning and direct preference optimization, with layer streaming technology that allows an 8 billion parameter model to train on a 4 GB laptop GPU. The tool integrates with popular frameworks like Hugging Face, PyTorch, and Ollama, and supports quantization formats such as GGUF and QLoRA for memory-efficient training.

Soup suits developers who want to fine-tune models locally without managing complex infrastructure or remote servers. It works well for projects where a single YAML configuration can capture all training parameters, and where hardware constraints make traditional fine-tuning setups impractical. The tool is designed for those seeking a streamlined alternative to manual configuration management in LLM training workflows.

The project maintains active continuous integration with passing tests. Development activity shows regular updates to the codebase and ongoing refinement of core functionality.