Introduction to Fine-Tuning Frameworks
CSGHub supports the following fine-tuning frameworks to meet different fine-tuning needs.
LLaMA-Factory
LLaMA-Factory is a unified and efficient fine-tuning framework that supports over 100 large language models (LLMs) and visual language models (VLMs). The framework provides various fine-tuning methods, including LoRA and QLoRA, which can significantly enhance training speed and efficiency. For instance, compared to ChatGLM's P-Tuning, LLaMA-Factory's LoRA fine-tuning achieves a 3.7 times increase in training speed for advertising copy generation tasks while also achieving better Rouge scores. Additionally, LLaMA-Factory supports 4-bit quantization techniques (QLoRA) to further optimize GPU memory utilization.
Links
MS-SWIFT
MS-SWIFT is a flexible and efficient fine-tuning framework that supports both PEFT (parameter-efficient fine-tuning) and full parameter fine-tuning methods. This framework is compatible with a variety of models and datasets, allowing users to choose the most suitable fine-tuning method based on their needs. MS-SWIFT offers rich features, including fine-tuning for embedding models, support for the GRPO (Group Relative Policy Optimization) algorithm, and integration with LMDeploy. Furthermore, MS-SWIFT provides detailed documentation and examples to help users get started quickly and resolve common issues.
Links