🏠 Homepage | 📜 Paper | 🤗 Dataset | 🚀 Installation

• [2025.06.03] 📑 Our paper is now available on arXiv.
• [2025.06.02] 🎉 We have released our Reasoning-Table Dataset on Hugging Face!
• [2025.06.02] 🎉 We have released the code of Reasoning-Table.

We release the full dataset used in the Reasoning-Table project for table reasoning tasks. The dataset is hosted on Hugging Face Datasets for easy access and download.

🔗 Dataset Link: TableQAKit/Reasoning-Table

You can download all files programmatically using the huggingface_hub library as shown below:

from huggingface_hub import hf_hub_download, login, list_repo_files
import os

repo_id = "TableQAKit/Reasoning-Table"

# Local download directory
download_dir = "you download path"
os.makedirs(download_dir, exist_ok=True)

# List and download all files from the dataset
all_files = list_repo_files(repo_id=repo_id, repo_type="dataset")

for file in all_files:
    print(f"Downloading: {file}")
    file_path = hf_hub_download(
        repo_id=repo_id,
        filename=file,
        repo_type="dataset",
        local_dir=download_dir,
        local_dir_use_symlinks=False
    )
    print(f"Saved to: {file_path}")

Please install torch, vllm and ray according to your own environment configuration. We provide a configuration example adapted from TinyZero in the following:

pip install -r requirements.txt

# install torch
pip install torch==2.4.0 --index-url https://download.pytorch.org/whl/cu121
# install vllm
pip install vllm==0.6.3
pip install ray

Please further install the verl in the current project and flash attention.

# verl
pip install -e .

# flash attention 2
pip install flash-attn --no-build-isolation
# quality of life
pip install wandb IPython matplotlib

For GRPO and PPO training, we provide configuration scripts for different model sizes. You can select the appropriate script based on your hardware resources:

# For Qwen2.5-3B-Instruct model training
bash train_grpo_3b.sh

# For Qwen2.5-7B-Instruct model training (requires more GPU memory)
bash train_grpo_7b.sh

Key parameters you may want to customize:

• BASE_MODEL: Path to the base model (e.g., "/Qwen/Qwen2.5-3B-Instruct")
• DATA_DIR: Path to your training dataset (e.g., data/wikitq)
• EXPERIMENT_NAME: Name for your training experiment (used for logging and checkpoints)
• trainer.n_gpus_per_node: Number of GPUs to use per node (adjust based on hardware)
• trainer.total_epochs: Number of training epochs

The training scripts use veRL and support both GRPO (Generalized Return Policy Optimization) and standard PPO algorithms for reinforcement learning.

We provide a powerful and easy-to-use evaluation script for table reasoning tasks. It supports multiple datasets and evaluation modes, making it ideal for benchmarking various models across different reasoning challenges.

To evaluate a trained model on a specific benchmark:

bash evaluation/test.sh

The test.sh script supports the following features:

• Multiple Tasks: Configurable for various benchmarks like WikiTableQuestions (wikitq), ToTTo (totto), and others via the TASK_NAME parameter.

• Evaluation Modes:

  • standard: Exact match evaluation (default)
  • llm: Uses LLM judge to evaluate responses (helpful for more complex reasoning tasks)
  • combined: Uses both exact match and LLM-based evaluation

• Model Configuration:

  • MODEL_PATH: Path to your fine-tuned model for inference
  • EVAL_MODEL_PATH: Path to evaluation model (when using LLM-based evaluation)
  • MODEL_SIZE: Size of your model (e.g., 3b, 7b)
  • TRAIN_TYPE: Training method (e.g., grpo, ppo, sft, base)

• Hardware Configuration:

  • TENSOR_PARALLEL_SIZE: Set based on your GPU configuration
  • BATCH_SIZE: Adjust based on available memory
  • LLM_EVAL_BATCH_SIZE: Batch size for LLM-based evaluation

For detailed usage instructions, parameter settings, and example scripts, please refer to the detailed evaluation README.

Our code is built upon veRL and TinyZero.

@misc{lei2025reasoningtableexploringreinforcementlearning,
      title={Reasoning-Table: Exploring Reinforcement Learning for Table Reasoning}, 
      author={Fangyu Lei and Jinxiang Meng and Yiming Huang and Tinghong Chen and Yun Zhang and Shizhu He and Jun Zhao and Kang Liu},
      year={2025},
      eprint={2506.01710},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.01710}, 
}