TreeAILab/NumericBench

Introduction

NumericBench is a comprehensive benchmark designed to evaluate the numerical reasoning capabilities of Large Language Models, addressing their limitations in tasks like arithmetic, number recognition, contextual retrieval, comparison, summarization, and logical reasoning. By incorporating diverse datasets ranging from synthetic number lists to real-world domains like stock trends and weather patterns, NumericBench systematically tests LLMs in both structured and noisy contexts. Experiments on models such as GPT-4o and DeepSeek-V3 reveal significant weaknesses, emphasizing the need for numerically-aware modeling to enhance LLMs' real-world applicability.

Github Repo: https://github.com/TreeAI-Lab/NumericBench

Arxiv Paper: https://arxiv.org/abs/2502.11075

Arxiv Paper on Hugging Face: https://huggingface.co/papers/2502.11075

How to use it?

Loading Data

from huggingface_hub import hf_hub_download
import json

dataset_name_list = ['arithmetic_operation/context_arithmetic_operation.json', 'arithmetic_operation/arithmetic_operation.json',
            'mixed_number_sting/mixed_number_string_500_per_sample.json', 
            'num_list/num_list_500_per_sample_1000_length.json', 'num_list/num_list_500_per_sample_100_length.json', 
            'sequence/sequence_500_sample_100_length.json', 
            'stock/single-turn/stock_500_per_sample_150_length.json', 'stock/single-turn/stock_500_per_sample_300_length.json', 'stock/multi-turn/stock_multi_turn_100_per_sample_100_length.json',
            'weather/single-turn/weather_500_per_sample_200_length.json', 'weather/single-turn/weather_500_per_sample_400_length.json', 'weather/multi-turn/weather_multi_turn_100_per_sample_100_length.json'
            ]

REPO_ID = "TreeAILab/NumericBench"

for dataset_name in dataset_name_list:
    with open(hf_hub_download(repo_id=REPO_ID, filename=dataset_name, repo_type="dataset")) as f:
        dataset = json.load(f)

Alternatively, you can download the dataset from this link.

Data Format

Due to the excessive length of the content, "..." is used to indicate omission.

single-turn

{
    "system_prompt": "...",
    "system_prompt_cot": "...",
    "description": "...",
    "data": [
        {
            "idx": 0,
            "question_index": 0,
            "question": "What is the result of A + B? Please round the answer to two decimal places. ",
            "struct_data": "{'A': 6.755, 'B': -1.225}",
            "answer": 5.53,
            "ability": "1-digit integer with 3 decimal places"
        }, ...
    ]
}

multi-turn

{
    "system_prompt": "...",
    "system_prompt_cot": "...",
    "description": "...",
    "data": [
        {
            "idx": 0,
            "multi_turn_QA": [
                {
                    "turn_index": 0,
                    "question_index": 6,
                    "struct_data": "...",
                    "question": "...",
                    "answer": "F",
                    "ability": "contextual retrieval"
                }, ...
            ]
        }, ...
    ]
}

Dataset statistics

For more details, please refer to our paper.

Citation

@misc{li2025exposingnumeracygapsbenchmark,
      title={Exposing Numeracy Gaps: A Benchmark to Evaluate Fundamental Numerical Abilities in Large Language Models}, 
      author={Haoyang Li and Xuejia Chen and Zhanchao XU and Darian Li and Nicole Hu and Fei Teng and Yiming Li and Luyu Qiu and Chen Jason Zhang and Qing Li and Lei Chen},
      year={2025},
      eprint={2502.11075},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.11075}, 
}