LLaMA-Factory / ms-swift (ModelScope SWIFT)

Two popular open-source fine-tuning frameworks for LLMs and VLMs, usable by just editing a config file.

LLaMA-Factory is an open-source fine-tuning framework maintained by the developer hiyouga (Yaowei Zheng and others), with an accompanying paper presented as a system demonstration at ACL 2024; ms-swift is the equivalent framework from Alibaba's ModelScope team. Both package full-parameter fine-tuning, parameter-efficient methods like LoRA/QLoRA, and preference-alignment and reinforcement-learning methods like DPO and GRPO into a configuration file or command-line interface, and both support multimodal models such as Qwen-VL and InternVL. In embodied AI, they're commonly used to supervised-fine-tune a VLM — for example, training a model for embodied reasoning, spatial question-answering, or task planning — which can then serve as a VLA's backbone or its high-level planner.

ExampleWrite a YAML file for LLaMA-Factory and use your own annotated robot-scene question-answering data to LoRA-fine-tune Qwen2.5-VL.

Also called

LlamaFactory, SWIFT, ms-swift

Related

Fine-tuning · Low-Rank Adaptation (LoRA) · Supervised Fine-Tuning · Qwen-VL series (Qwen2.5-VL / Qwen3-VL) · ModelScope · Hugging Face Transformers

Sources

LLaMA-Factory GitHub
ms-swift GitHub

As of

2026-09