DuetDA

Official code for DuetDA: Decomposed and Dynamic Data Attribution with Model-State Gating for Accelerated Scientific Endeavors.

This repo supports:

• Building MatBench CGCNN splits (count-based OOD and difficulty-based OOD)
• Meta-training a DuetDA data valuator
• Training CGCNN/SchNet/ALIGNN with DuetDA-only data selection

We appreciate your citations:

@inproceedings{duetda2026,
author = {Chen, Jianpeng and Zhan, Wangzhi and Wang, Haohui and Fu, Dongqi and Zhou, Dawei},
title = {DuetDA:  Decomposed and Dynamic Data Attribution with Model-State Gating for Accelerated Scientific Endeavors},
year = {2026},
isbn = {9798400722592},
publisher = {Association for Computing Machinery},
doi = {10.1145/3770855.3818877},
booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2},
pages = {10589–10600},
numpages = {12},
series = {KDD '26}
}

1) Environment

Recommended: Python 3.10+.

python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install torch numpy pandas scikit-learn pymatgen matbench wandb pyarrow

2) Prepare data splits

2.1 Count-based OOD split

python prepare_cgcnn_count_based_ood.py

Output (default):

• cgcnn_data/matbench_log_kvrh_CountBasedOOD/

2.2 Difficulty-based OOD split

python prepare_cgcnn_difficulty.py \
	--fit_csv fit_difficulty.csv

Output (default):

• cgcnn_data/matbench_log_kvrh_difficultyOOD1/

3) Meta-train DuetDA valuator

python modules/meta_train_cgcnn.py \
	--data_root cgcnn_data/matbench_log_kvrh_CountBasedOOD \
	--checkpoint-dir checkpoints/meta_train/cgcnn \
	--backbone cgcnn \
	--fold 1 \
	--num-models 3 \
	--num-outer-steps 50 \
	--truncation-steps 3 \
	--inner-steps 5 \
	--meta-lr 0.001 \
	--batch-size 256

Example checkpoint to use later:

• checkpoints/meta_train/cgcnn/data_attributor_meta_step_*.pt

4) Train with DuetDA curation (quick run)

main.py now supports only --da-method duetda.

python main.py \
	--data-root cgcnn_data/matbench_log_kvrh_difficultyOOD1 \
	--data-name cgcnn_matbench \
	--da-model-ckpt checkpoints/meta_train/cgcnn/data_attributor_meta_step_50.pt \
	--model-name schnet \
	--task-type regression \
	--da-method duetda \
	--fold 1 \
	--epochs 2 \
	--cuda \
	--optim Adam \
	--batch-size 256 \
	--print-freq 5 \
	--seed 42 \
	--selection-ratio 0.5

Optional W&B logging:

--use-wandb --wandb-entity <entity> --wandb-project <project> --wandb-name <run_name>

5) Key outputs

• Best checkpoint: checkpoints/<data>_duetda_<ratio>_model_best.pth.tar
• Last checkpoint: checkpoints/<data>_duetda_<ratio>_last.pth.tar
• Test predictions: <duetda>_<ratio>_<data>_test_results.csv (or --test-res-path)

Notes

• Use split fold values consistent across data prep, meta-training, and final training.
• If you train on CPU, remove --cuda.