Soup

Why we included this project

Fine-tuning a modern LLM on a laptop or a single consumer GPU usually means juggling quantization, batch sizes, and adapter configs by hand. Soup collapses that into one YAML file and a single `soup train` command, so a small team can go from raw data to a tuned model without renting a GPU cluster. Its layer streaming keeps the frozen base model out of VRAM and feeds it through one decoder layer at a time, which is how an 8B model can train on a 4 GB card with QLoRA-style NF4 quantization. Beyond supervised fine-tuning it also handles preference methods like DPO, ORPO, SimPO, and KTO, plus utilities to synthesize reward functions, merge adapters, and export to GGUF for local serving. That makes it a practical choice for developers who want to own their model tuning on hardware they already have.