Install Packages:

pip install -r requirements.txt

ConsTRev Dataset Creation Pipeline:

Generate the ConsTRev's constraint instruction with: python put_constraint.py

TRIPS:

Build the planner:

1. Create synthetic trajectory using GPT-4o with: python create_syn_data_from_gpt4o.py
2. Build an initial planner: a. Go to finetune/train_script folder. b. Run bash run.sh script.
3. Build self-training data with: python create_self_training_data.py
4. Enhance the planner via self-training: a. Go to align/runs folder. b. Run bash llama3_instruct_cpo_simpo.sh script.
5. Iterate through steps 3 and 4.

Run Search:

1. Go to the MCTS folder.
2. Pass in the path for the planner and reviser in the run.sh file.
3. run bash run.sh script.

Test Data and Predictions:

1. Test data is in data/test_input folder.
2. Prediction data is in data/output folder.

Citation

If you found our paper or code useful, please cite as:

@inproceedings{cao-ng-2025-craft,
    title = "A Constrained Text Revision Agent via Iterative Planning and Searching",
    author = "Cao, Hannan  and
      Ng, Hwee Tou",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2025",
    year = "2025",
}