MakazhanAlpamys/Soup

Health v2 · maintenance only

Adoption not part of the score

3102 stars · 463 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

Soup is a Python CLI that fine-tunes and post-trains LLMs from a single YAML config, supporting 23 methods (SFT, DPO, ORPO, SimPO, KTO, etc.) with automatic config generation, data pre-flight checks, and eval gating. Its layer-streaming feature streams the frozen base model from RAM or NVMe one layer at a time, letting an 8B model fine-tune on a 4 GB consumer GPU.

Use cases

• fine-tune llama 3.1 8b on my 4gb laptop gpu
• run dpo training without renting a big gpu
• convert my llamafactory config to another tool
• sft a local llm from one yaml file
• export fine-tuned model to gguf for ollama
• train lora adapter on consumer gpu with low vram
• migrate axolotl config to a new finetuning cli
• evaluate whether my finetune actually improved before shipping

When to choose

• you want one-command LLM fine-tuning on consumer hardware with limited VRAM
• you prefer declarative YAML configs over writing training scripts
• you need preference alignment methods (DPO/ORPO/SimPO/KTO) on small GPUs
• you're migrating from LLaMA-Factory, Axolotl, or Unsloth and want automatic config conversion

When to avoid

• you need large-scale multi-node distributed training on clusters
• you need RLHF methods like PPO or GRPO with generation rollouts on streamed bases
• you require guaranteed production-grade correctness on bleeding-edge releases - recent changelogs document silent correctness defects in some paths
• you want a GUI workbench - the desktop app is still upcoming

Facets

cli-tool · maturity active

llm-training machine-learning cli configuration-management testing large-language-models machine-learning developer-tools python cli windows lora qlora peft gguf ollama huggingface dpo sft low-vram layer-streaming consumer-gpu local-llm llmops config-migration command-line macos linux gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="MakazhanAlpamys/Soup")