MMSci_Table

Dataset for the paper "Does Table Source Matter? Benchmarking and Improving Multimodal Scientific Table Understanding and Reasoning"

📑 Paper    Github

MMSci Dataset Collection

The MMSci dataset collection consists of three complementary datasets designed for scientific multimodal table understanding and reasoning: MMSci-Pre, MMSci-Ins, and MMSci-Eval.

Dataset Summary

• MMSci-Pre: A domain-specific pre-training dataset containing 52K scientific table structure recognition samples
• MMSci-Ins: An instruction tuning dataset with 12K samples across three table-based tasks
• MMSci-Eval: A benchmark with 3,114 testing samples for numerical reasoning evaluation

Framework Overview

Figure 1: Overview of the MMSci framework showing the four key stages: Table Image Generation, Dataset Construction, Table Structure Learning, and Visual Instruction Tuning.

Dataset Details

MMSci-Pre

• Size: 52K samples
• Format: Table image-to-HTML pairs
• Source: Scientific papers from SciGen dataset
• Purpose: Table structure learning and alignment of visual features with textual representations
• Features:
  • High-quality HTML format tables
  • Rendered table images preserving structural integrity
  • Complex layouts and relationships from scientific papers
  • Focus on tables with significant numerical values

Figure 2: Example from MMSci-Pre dataset showing the table image and its corresponding HTML representation.

MMSci-Ins

• Size: 12K samples
• Format: Instruction-following samples with reasoning steps
• Tasks:
  • Table Question Answering (TQA)
  • Table Fact Verification (TFV)
  • Table-to-Text Generation (T2T)
• Features:
  • Detailed step-by-step reasoning processes
  • Balanced distribution across three tasks
  • Each table paired with one TQA, TFV, and T2T task
  • Built upon scientific domain tables

Figure 3: Example from MMSci-Ins dataset showing instruction-following samples across different tasks.

MMSci-Eval

• Size: 3,114 samples
• Purpose: Comprehensive evaluation of numerical reasoning capabilities
• Features:
  • Testing samples across TQA, TFV, and T2T tasks
  • Focus on numerical reasoning assessment
  • Based on SciGen dataset test set
  • Diverse reasoning types and complexity levels

Dataset Creation

The datasets were created through a rigorous process:

1. Collection of raw tabular data from SciGen dataset
2. Transformation of textual tables into HTML format
3. Rendering of HTML tables into high-quality images
4. Generation of instruction-following samples with reasoning steps
5. Quality assurance through balanced task distribution

Intended Uses

• Pre-training multimodal language models for table understanding
• Fine-tuning models for specific table-based tasks
• Evaluating numerical reasoning capabilities in scientific contexts
• Benchmarking table understanding and reasoning systems

Table Question Answering (TQA)

Figure 4: Example of a TQA task showing the question, reasoning steps, and answer.

Table Fact Verification (TFV)

Figure 5: Example of a TFV task showing the statement, verification process, and conclusion.

Citation

If you found this repository or paper is helpful to you, please cite our paper.

@misc{yang2025doestablesourcematter,
      title={Does Table Source Matter? Benchmarking and Improving Multimodal Scientific Table Understanding and Reasoning}, 
      author={Bohao Yang and Yingji Zhang and Dong Liu and André Freitas and Chenghua Lin},
      year={2025},
      eprint={2501.13042},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2501.13042}, 
}

}