ChatGLM-Efficient-Tuning icon indicating copy to clipboard operation
ChatGLM-Efficient-Tuning copied to clipboard

Fine-tuning ChatGLM-6B with PEFT | 基于 PEFT 的高效 ChatGLM 微调

ChatGLM Efficient Tuning

GitHub Repo stars GitHub Code License GitHub last commit PyPI GitHub pull request

Fine-tuning 🤖ChatGLM-6B model with 🤗PEFT.

👋 Join our WeChat.

[ English | 中文 ]


[23/07/15] Now we develop an all-in-one Web UI for training, evaluation and inference. Try to fine-tune ChatGLM-6B model in your Web browser. Thank @KanadeSiina for his efforts in the development.

[23/07/09] Now we release FastEdit⚡🩹, an easy-to-use package for editing the factual knowledge of large language models efficiently. Please follow FastEdit if you are interested.

[23/06/25] Now we align the demo API with the OpenAI's format where you can insert the fine-tuned model in arbitrary ChatGPT-based applications.

[23/06/25] Now we support fine-tuning the ChatGLM2-6B model with our framework!

[23/06/05] Now we support 4-bit LoRA training (aka QLoRA). Try --quantization_bit 4 argument to work with 4-bit quantized model. (experimental feature)

[23/06/01] We implemented a framework supporting the efficient tuning of LLaMA and BLOOM models. Please follow LLaMA-Efficient-Tuning if you are interested.

[23/05/19] Now we support using the development set to evaluate the model while training. Try --dev_ratio argument to specify the size of development set.

[23/04/29] Now we support training ChatGLM with Reinforcement Learning with Human Feedback (RLHF) ! We provide several examples to run RLHF training, please refer to the examples folder for details.

[23/04/20] Our repo achieved 100 stars within 12 days! Congratulations!

[23/04/19] Now we support merging the weights of fine-tuned models trained by LoRA! Try --checkpoint_dir checkpoint1,checkpoint2 argument for continually fine-tuning the models.

[23/04/18] Now we support training the quantized models using three fine-tuning methods! Try quantization_bit argument for training the model in 4/8 bits.

[23/04/12] Now we support training from checkpoints! Use --checkpoint_dir argument to specify the checkpoint model to fine-tune from.

[23/04/11] Now we support training with combined datasets! Try --dataset dataset1,dataset2 argument for training with multiple datasets.


Our script now supports the following datasets:

Please refer to data/ for details.

Some datasets require confirmation before using them, so we recommend logging in with your HuggingFace account using these commands.

pip install --upgrade huggingface_hub
huggingface-cli login

Fine-Tuning Methods

Our script now supports the following fine-tuning methods:

  • LoRA
    • Fine-tuning the low-rank adapters of the model.
  • P-Tuning V2
    • Fine-tuning the prefix encoder of the model.
  • Freeze
    • Fine-tuning the MLPs in the last n blocks of the model.
  • Full Tuning
    • Fine-tuning all the parameters of the model.


  • Python 3.8+ and PyTorch 1.13.1+
  • 🤗Transformers, Datasets, Accelerate, PEFT and TRL
  • fire, protobuf, cpm-kernels and sentencepiece
  • jieba, rouge-chinese and nltk (used at evaluation)
  • gradio and matplotlib (used in
  • uvicorn, fastapi and sse-starlette (used in

And powerful GPUs!

Getting Started

Data Preparation (optional)

Please refer to data/example_dataset for checking the details about the format of dataset files. You can either use a single .json file or a dataset loading script with multiple files to create a custom dataset.

Note: please update data/dataset_info.json to use your custom dataset. About the format of this file, please refer to data/

Dependence Installation (optional)

git clone
conda create -n chatglm_etuning python=3.10
conda activate chatglm_etuning
cd ChatGLM-Efficient-Tuning
pip install -r requirements.txt

If you want to enable LoRA(QLoRA) or Freeze quantization on Windows, you will be required to install a pre-built version of bitsandbytes library, which supports CUDA 11.1 to 12.1.

pip install

Fine-tuning with a Single GPU

CUDA_VISIBLE_DEVICES=0 python src/ \
    --stage sft \
    --model_name_or_path path_to_your_chatglm_model \
    --do_train \
    --dataset alpaca_gpt4_en \
    --finetuning_type lora \
    --output_dir path_to_sft_checkpoint \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 5e-5 \
    --num_train_epochs 3.0 \

Please refer to our Wiki about the details of the arguments.

Distributed Fine-tuning with Multiple GPUs

accelerate config # configure the environment
accelerate launch src/ # arguments (same as above)

Training Reward Model

CUDA_VISIBLE_DEVICES=0 python src/ \
    --stage rm \
    --model_name_or_path path_to_your_chatglm_model \
    --do_train \
    --dataset comparison_gpt4_en \
    --finetuning_type lora \
    --output_dir path_to_rm_checkpoint \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 1e-5 \
    --num_train_epochs 1.0 \

Training with RLHF

CUDA_VISIBLE_DEVICES=0 python src/ \
    --stage ppo \
    --model_name_or_path path_to_your_chatglm_model \
    --do_train \
    --dataset alpaca_gpt4_en \
    --finetuning_type lora \
    --resume_lora_training False \
    --checkpoint_dir path_to_sft_checkpoint \
    --reward_model path_to_rm_checkpoint \
    --output_dir path_to_ppo_checkpoint \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 1e-5 \
    --num_train_epochs 1.0 \

Evaluation (BLEU and ROUGE_CHINESE)

CUDA_VISIBLE_DEVICES=0 python src/ \
    --stage sft \
    --model_name_or_path path_to_your_chatglm_model \
    --do_eval \
    --dataset alpaca_gpt4_en \
    --checkpoint_dir path_to_checkpoint \
    --output_dir path_to_eval_result \
    --per_device_eval_batch_size 8 \
    --max_samples 50 \


CUDA_VISIBLE_DEVICES=0 python src/ \
    --stage sft \
    --model_name_or_path path_to_your_chatglm_model \
    --do_predict \
    --dataset alpaca_gpt4_en \
    --checkpoint_dir path_to_checkpoint \
    --output_dir path_to_predict_result \
    --per_device_eval_batch_size 8 \
    --max_samples 50 \

API / CLI Demo

python src/ \
    --model_name_or_path path_to_your_chatglm_model \
    --checkpoint_dir path_to_checkpoint

All-in-one Web UI

python src/

Export model

python src/ \
    --model_name_or_path path_to_your_chatglm_model \
    --checkpoint_dir path_to_checkpoint \
    --output_dir path_to_export

Hardware Requirements

Fine-tune method Batch size Mode GRAM Speed
LoRA (r=8) 16 FP16 28GB 8ex/s
LoRA (r=8) 8 FP16 24GB 8ex/s
LoRA (r=8) 4 FP16 20GB 8ex/s
LoRA (r=8) 4 INT8 10GB 8ex/s
LoRA (r=8) 4 INT4 8GB 8ex/s
P-Tuning (p=16) 4 FP16 20GB 8ex/s
P-Tuning (p=16) 4 INT8 16GB 8ex/s
P-Tuning (p=16) 4 INT4 12GB 8ex/s
Freeze (l=3) 4 FP16 24GB 8ex/s
Freeze (l=3) 4 INT8 12GB 8ex/s
RM method Batch size Mode GRAM Speed
LoRA (r=8) + rm 4 FP16 22GB -
LoRA (r=8) + rm 1 INT8 11GB -
RLHF method Batch size Mode GRAM Speed
LoRA (r=8) + ppo 4 FP16 23GB -
LoRA (r=8) + ppo 1 INT8 12GB -

Note: r is the lora rank, p is the number of prefix tokens, l is the number of trainable layers, ex/s is the examples per second at training. The gradient_accumulation_steps is set to 1. All are evaluated on a single Tesla V100 (32G) GPU, they are approximated values and may vary in different GPUs.

Fine-tuning ChatGLM: A Case

Training Results

We use the whole alpaca_gpt4_zh dataset to fine-tune the ChatGLM model with LoRA (r=8) for one epoch, using the default hyper-parameters. The loss curve during training is presented below.

training loss

Evaluation Results

We select 100 instances in the alpaca_gpt4_zh dataset to evaluate the fine-tuned ChatGLM model and compute the BLEU and ROUGE scores. The results are presented below.

Score Original FZ (l=2) PT (p=16) LoRA (r=8)
BLEU-4 15.75 16.85 16.06 17.01 (+1.26)
Rouge-1 34.51 36.62 34.80 36.77 (+2.26)
Rouge-2 15.11 17.04 15.32 16.83 (+1.72)
Rouge-l 26.18 28.17 26.35 28.86 (+2.68)
Params (%) / 4.35% 0.06% 0.06%

FZ: freeze tuning, PT: P-Tuning V2 (we use pre_seq_len=16 for fair comparison with LoRA), Params: the percentange of trainable parameters.

Compared with Existing Implementations

  • THUDM/ChatGLM-6B
    • Official implementation of fine-tuning ChatGLM with P-Tuning v2 on the ADGEN dataset.
    • Our fine-tuning script is largely depend on it. We further implement the LoRA tuning method. Additionally, we dynamically pad the inputs to the longest sequence in the batch instead of the maximum length, to accelerate the fine-tuning.
  • mymusise/ChatGLM-Tuning
    • An unoffical implementation of fine-tuning ChatGLM with LoRA on the Stanford Alpaca dataset.
    • We borrowed some ideas from it. Our fine-tuning script integrates the data pre-processing part into the training procedure, so we need not generate a pre-processed dataset before training.
  • ssbuild/chatglm_finetuning
  • lich99/ChatGLM-finetune-LoRA
  • liucongg/ChatGLM-Finetuning
    • An unofficial implementation of fine-tuning ChatGLM with several methods including Freeze, LoRA and P-Tuning on the industrial dataset.
    • We are aim to incorporate more instruction-following datasets for fine-tuning the ChatGLM model.
  • yanqiangmiffy/InstructGLM
    • An unofficial implementation of fine-tuning ChatGLM that explores the ChatGLM's ability on the instruction-following datasets.
    • Our fine-tuning script integrates the data pre-processing part in to the training procedure.


  • [ ] Employing LangChain to easily build applications that are capable of leveraging external knowledge upon fine-tuned ChatGLM models.
  • [ ] Implementing the alignment algorithms to align human preferrences.
  • [ ] Incorporating Chinese datasets into the training sets.
  • [ ] Incorporating ChatGPT & GPT-4 self-chat data into the training sets.
  • [x] Implementing the Freeze-Tuning and P-Tuning method.
  • [x] Supporting Multi-GPUs fine-tuning.
  • [x] Adding script for evaluation.
  • [x] Loading from checkpoint.
  • [x] Fine-tuning the quantized model.
  • [x] Writing a guidebook about how to fine-tune ChatGLM with this framework.
  • [ ] Combining with state-of-the-art model editing algorithms. (e.g. MEND)
  • [x] Incorporating the OpenAssistant Conversations Dataset for SFT and alignment.
  • [ ] Incorporating the high quality Chinese instruction dataset COIG.


This repository is licensed under the Apache-2.0 License. Please follow the Model License to use ChatGLM-6B model.


If this work is helpful, please cite as:

  title = {ChatGLM Efficient Tuning},
  author = {hiyouga},
  howpublished = {\url{}},
  year = {2023}


This repo benefits from ChatGLM-6B, ChatGLM-Tuning and yuanzhoulvpi2017/zero_nlp. Thanks for their wonderful works.

Star History

Star History Chart