This repository provides datasets, models and code for CoEdIT the instruction-tuned text editing models, with the official implementation of the following paper:

CoEdIT: Text Editing by Task-Specific Instruction Tuning
Vipul Raheja, Dhruv Kumar, Ryan Koo, and Dongyeop Kang

Our code is based on Hugging Face transformers.

1. Clone the repository

   git clone https://github.com/vipulraheja/coedit.git
   

2. Run the setup script

   $ cd coedit
   $ sh setup_env.sh
   

Available on Hugging Face. Example data point:

{
  '_id': 1,
  'task': "gec",
  'src': "Improve the grammaticality: As the number of people grows, the need of habitable environment is unquestionably essential.",
  'tgt': "As the number of people grows, the need for a habitable environment is unquestionably increasing."
}

Please note that this dataset contains 69k instances (as opposed to the 82k instances we used in the paper). This is because this public release includes only the instances that were acquired and curated from publicly available datasets. Specifically, it is missing roughly 13k instances in training and 1.5k instances in validation data from Simplification and Formality Transfer tasks due to licensing restrictions.

Example script for the CoEdIT-xl model.

sh train/train_coedit_xl.sh

The CoEdIT-large and CoEdIT-xxl models can be trained by making the corresponding changes to this script.

We have uploaded all our model checkpoints to Hugging Face.

You can directly load our models using Hugging Face Transformers.

from transformers import AutoTokenizer, T5ForConditionalGeneration

tokenizer = AutoTokenizer.from_pretrained("grammarly/coedit-xl")
model = T5ForConditionalGeneration.from_pretrained("grammarly/coedit-xl")
input_text = 'Fix grammatical errors in this sentence: When I grow up, I start to understand what he said is quite right.'
input_ids = tokenizer(input_text, return_tensors="pt").input_ids
outputs = model.generate(input_ids, max_length=256)
edited_text = tokenizer.decode(outputs[0], skip_special_tokens=True)

@article{raheja2023coedit,
      title={CoEdIT: Text Editing by Task-Specific Instruction Tuning}, 
      author={Vipul Raheja and Dhruv Kumar and Ryan Koo and Dongyeop Kang},
      url={https://arxiv.org/abs/2305.09857},
      year={2023},
      eprint={2305.09857},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}