ChatGLM-Finetuning

Why we included this project

This repository collects working examples for adapting ChatGLM-6B, ChatGLM2-6B, and ChatGLM3-6B to your own tasks through four fine-tuning approaches: freeze (keeping most layers fixed), LoRA, P-Tuning, and full-parameter updates. Training runs on DeepSpeed directly rather than wrapping the standard Trainer, so the code is easier to trace and change when you need finer control over the loop. You get single- and multi-GPU launch scripts plus gradient checkpointing and offload flags to fit models into limited VRAM. It also includes evaluation support for information extraction, generation, and classification tasks. If you work with Chinese-language NLP or want a readable reference for fine-tuning a 6B-scale model on modest hardware, this is a good starting point, though it is training code rather than a deployment service.