The latest release of hiyouga/llamafactory is v0.9.5. This repository is a Python tool for fine-tuning 100+ large language models (LLMs) and vision-language models (VLMs) in one place. You set up the fine-tuning workflow through a command-line interface (CLI) and a web user interface, without writing any code. However, the changed items in v0.9.5 are not broken down in this record.
Fine-tuning is the work of training a pretrained model again to fit specific data. This project provides full training, which learns all weights, LoRA, which learns only the added weights, its quantized version QLoRA, and alignment training, all together through the CLI and the GUI. Covering models and training methods within a single tool is the essence of this repository. The README explains that 100+ LLMs can be fine-tuned from the command line and the web interface without writing code. The license is Apache-2.0, and the repository description also lists ACL 2024. The star count is 75627, and the repository was created on May 28, 2023.
That said, this record contains no in-house benchmark showing training speed or stability. The performance score and the quality score are unfilled, and only the momentum score remains, at 5.5 out of 10. It is difficult to gauge the performance or quality of v0.9.5 from the provided information alone. The sensible basis for any judgment is to check the release notes and the actual training results separately.