Environment Setup

To set up the environment, ensure you have the following versions installed:

• CUDA: 12.1
• Python: 3.10

To install the required Python dependencies, run the following command:

pip install -r requirements.txt

Download data：

English Data

• BEA-2019 Shared Task: Download Here
• CoNLL-2014 Shared Task: Download Here

Chinese Data

• FCGEC Dataset: Download Here
• NaCGEC Dataset: Download Here

Data Preprocessing

You need to preprocess the datasets into a unified format, compatible with the training pipeline.

Required Format

Ensure that the data is structured in the same format as the provided examples:

• data/epo_data_sample.json
• data/sft_data_example.json

Additionally, you will need to modify the data/dataset_info.json file to match the specifics of your dataset configuration.

Training Pipeline

SFT Stage 1:

bash bash/train_gec_sft_stage1.sh

bash bash/export_model.sh  # merge lora weight

SFT Stage 2:

bash bash/train_gec_sft_stage2.sh

Sampling

bash bash/gec_pairwise_sampling.sh  # generate pairwise samples

EPO Training

bash bash/train_gec_epo.sh

Note: For Chinese GEC, you can find the corresponding scripts in the bash directory.

Evaluation

bash bash/gec_eval.sh  # for English GEC model

bash bash/cgec_eval.sh  # for Chinese GEC model

Acknowledgements

This project is built upon LLaMA-Factory and utilizes the following tools for evaluation:

• ERRANT
• ChERRANT
• M2Scorer

We are grateful for their contributions.