Soup turns the pain of LLM fine‑tuning into a simple workflow. One config, one command, done. Install with pip install "soup‑cli[train]" and you’re ready to run soup init --template chatsoup and soup train --config soup.yaml. The tool eliminates SSH, config hell, and infrastructure headaches—training runs locally on your GPU with QLo RA, no cloud needed.

Version v0.72.4 adds layer streaming for DPO, ORPO, Sim PO, and KTO, keeping the frozen base out of VRAM and feeding it one decoder layer at a time. On an RTX 3050 4 GB the streamed DPO peaks at 0.914× the supervised‑fine‑tuning peak, while a second model would cost an extra 730 MB. Quantization (4‑bit NF4) allows an 8 B model on a 4 GB card, and batch sizes auto‑adjust.

Soup also ships a reward synth that infers a deterministic verifier from JSONL data, and a ship gate that emits evidence‑bound verdicts. Commands like soup ship --push owner/repo post strikt SHIP/DON’T‑SHIP cards. All features are documented in the full reference under docs/.