Official Code Implementation of TaLK

• Title: TaLK: Text-attributed Graph Dataset Distillation via Coupling Language Model with Graph-Aware Kernel
• Venue: EMNLP 2026 Findings
• Description: Official implementation of the dataset distillation algorithm for text-attributed graphs (TAGs) under joint LM-GNN training.
• Authors: Yeongho Kim, Yeonje Choi, and Kijung Shin
• Affiliation: KAIST AI
• Paper Link: https://arxiv.org/abs/2606.22975

The statistics of the datasets used in this work are summarized below:

• Cora / Arxiv: Citation network datasets.
  • The raw text and graph structures are sourced from here.
• Photo / Computers: Amazon co-purchase datasets.
  • The raw text and graph structures are sourced from here.

Note on Data Preparation: To download and set up the datasets, please follow the respective instructions provided in the source repositories linked above.

To distill a text-attributed graph dataset, run the following command:

python distillation.py --dataset [cora, Photo, Computers, arxiv] --data_root [ROOT FOLDER] --lm_name [bert, roberta, deberta] --seed [SEED] --cond_size [COND_SIZE] --gpu_id [GPU_ID]

To train and evaluate models using the generated synthetic (distilled) dataset, use:

python evaluation.py --dataset [cora, Photo, Computers, arxiv] --data_root [ROOT FOLDER] --lm_name [bert, roberta, deberta] --seed [SEED] --cond_size [COND_SIZE] --gpu_id [GPU_ID] --cond_path [PATH FOR THE SAVED SYNTHETIC DATA]

If you use this code in your research, please cite:

@article{kim2026talk,
  title={TaLK: Text-attributed Graph Dataset Distillation via Coupling Language Model with Graph-Aware Kernel},
  author={Kim, Yeongho and Choi, Yeonje and Shin, Kijung},
  journal={arXiv preprint arXiv:2606.22975},
  year={2026}
}