README.md

metadata

license: cc-by-4.0
task_categories:
  - text-generation
language:
  - en
tags:
  - graph-reasoning
  - llm-agent
  - curriculum-learning

GraphDancer Dataset

This repository contains the official training and evaluation data for GraphDancer, a framework for training Large Language Models (LLMs) to explore and reason over heterogeneous graphs via two-stage curriculum post-training.

• Paper: GraphDancer: Training LLMs to Explore and Reason over Graphs via Two-Stage Curriculum Post-Training
• Project Page: https://yuyangbai.com/graphdancer/
• Repository: https://github.com/leopoldwhite/GraphDancer

Dataset Summary

The dataset provides multi-turn trajectories for graph-based reasoning tasks. It includes:

• Training Data: Academic domain data used for both Curriculum-PPO and Curriculum-DPO stages.
• Test Data: Evaluation sets across four domains from GRBench: Biomedical (Healthcare), Goodreads (Literature), Amazon (E-commerce), and Legal.

Sample Usage

You can download the training and test data using the huggingface_hub library with the following script:

from huggingface_hub import hf_hub_download
import os, shutil

repo_id = "yuyangbai/GraphDancer-data"

os.makedirs("data/train", exist_ok=True)
for domain in ["biomedical", "goodreads", "amazon", "legal"]:
    os.makedirs(f"data/test/{domain}", exist_ok=True)

# Download training data
train_file = hf_hub_download(repo_id=repo_id, filename="train/train.parquet", repo_type="dataset")
shutil.copy2(train_file, "data/train/train.parquet")

# Download test data for each domain
for domain in ["biomedical", "goodreads", "amazon", "legal"]:
    test_file = hf_hub_download(repo_id=repo_id, filename=f"test/{domain}/test.parquet", repo_type="dataset")
    shutil.copy2(test_file, f"data/test/{domain}/test.parquet")

Structure

The expected directory structure for the data is as follows:

data/
├── train/train.parquet          # Academic training set
└── test/<domain>/test.parquet   # Per-domain test sets

Citation

@misc{bai2026graphdancertrainingllmsexplore,
      title={GraphDancer: Training LLMs to Explore and Reason over Graphs via Two-Stage Curriculum Post-Training},
      author={Yuyang Bai and Zhuofeng Li and Ping Nie and Yu Wang and Jianwen Xie and Yu Zhang},
      year={2026},
      eprint={2602.02518},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.02518},
}