ChatGLM Fine-Tuning, Soup-to-Nuts, With an Expiry Date

What it does

Provides a unified training framework for ChatGLM-6B and ChatGLM2-6B using parameter-efficient methods like LoRA, P-Tuning V2, and full fine-tuning. It bundles supervised fine-tuning, reward-model training, and RLHF into one CLI and Web UI, then serves the result through an OpenAI-compatible API, a Gradio web demo, or a local CLI.

The interesting bit

The project treats the whole lifecycle—training, evaluation, inference, and export—as a single pipeline rather than a collection of scripts. It also supports 4-bit quantized LoRA (QLoRA) and merging multiple LoRA checkpoints for continual fine-tuning, which keeps VRAM requirements merely large instead of astronomical.

Key highlights

• Supports LoRA, P-Tuning V2, Freeze, and full-parameter fine-tuning.
• Includes RLHF training (reward modeling + PPO) out of the box.
• Offers an all-in-one Web UI for training, evaluation, and inference (single-GPU only).
• Exposes an OpenAI-format API for dropping fine-tuned models into existing ChatGPT clients.
• Bundles dataset loaders for Alpaca, BELLE, ShareGPT, UltraChat, and others.

Caveats

• The repository is explicitly no longer maintained; the authors direct users to LLaMA-Factory for future work, including ChatGLM2-6B support.
• The Web UI is limited to single-GPU training.
• QLoRA support is marked as experimental.

Verdict

Worth a look if you are specifically targeting ChatGLM-6B and want a batteries-included tuner, but new projects should probably start with LLaMA-Factory instead.