ScienceMetaBench

English | 中文

🤗 HuggingFace Dataset | 💻 GitHub Repository

ScienceMetaBench is a benchmark dataset for evaluating the accuracy of metadata extraction from scientific literature PDF files. The dataset covers three major categories: academic papers, textbooks, and ebooks, and can be used to assess the performance of Large Language Models (LLMs) or other information extraction systems.

📊 Dataset Overview

Data Types

This benchmark includes three types of scientific literature:

1. Papers

   • Mainly from academic journals and conferences
   • Contains academic metadata such as DOI, keywords, etc.

2. Textbooks

   • Formally published textbooks
   • Includes ISBN, publisher, and other publication information

3. Ebooks

   • Digitized historical documents and books
   • Covers multiple languages and disciplines

Data Batches

This benchmark has undergone two rounds of data expansion, with each round adding new sample data:

data/
├── 20250806/          # First batch (August 6, 2024)
│   ├── ebook_0806.jsonl
│   ├── paper_0806.jsonl
│   └── textbook_0806.jsonl
└── 20251022/          # Second batch (October 22, 2024)
    ├── ebook_1022.jsonl
    ├── paper_1022.jsonl
    └── textbook_1022.jsonl

Note: The two batches of data complement each other to form a complete benchmark dataset. You can choose to use a single batch or merge them as needed.

PDF Files

The pdf/ directory contains the original PDF files corresponding to the benchmark data, with a directory structure consistent with the data/ directory.

File Naming Convention: All PDF files are named using their SHA256 hash values, in the format {sha256}.pdf. This naming scheme ensures file uniqueness and traceability, making it easy to locate the corresponding source file using the sha256 field in the JSONL data.

📝 Data Format

All data files are in JSONL format (one JSON object per line).

Academic Paper Fields

{
  "sha256": "SHA256 hash of the file",
  "origin_path": "Original path of the PDF file",
  "doi": "Digital Object Identifier",
  "title": "Paper title",
  "author": "Author name",
  "keyword": "Keywords (comma-separated)",
  "abstract": "Abstract content",
  "pub_time": "Publication year"
}

Textbook/Ebook Fields

{
  "sha256": "SHA256 hash of the file",
  "origin_path": "Original path of the PDF file",
  "isbn": "International Standard Book Number",
  "title": "Book title",
  "author": "Author name",
  "abstract": "Introduction/abstract",
  "category": "Classification number (e.g., Chinese Library Classification)",
  "pub_time": "Publication year",
  "publisher": "Publisher"
}

📖 Data Examples

Academic Paper Example

The following image shows an example of metadata fields extracted from an academic paper PDF:

As shown in the image, the following key information needs to be extracted from the paper's first page:

• DOI: Digital Object Identifier (e.g., 10.1186/s41038-017-0090-z)
• Title: Paper title
• Author: Author name
• Keyword: List of keywords
• Abstract: Paper abstract
• pub_time: Publication time (usually the year)

Textbook/Ebook Example

The following image shows an example of metadata fields extracted from the copyright page of a Chinese ebook PDF:

As shown in the image, the following key information needs to be extracted from the book's copyright page:

• ISBN: International Standard Book Number (e.g., 978-7-5385-8594-0)
• Title: Book title
• Author: Author/editor name
• Publisher: Publisher name
• pub_time: Publication time (year)
• Category: Book classification number
• Abstract: Content introduction (if available)

These examples demonstrate the core task of the benchmark test: accurately extracting structured metadata information from PDF documents in various formats and languages.

📊 Evaluation Metrics

Core Evaluation Metrics

This benchmark uses a string similarity-based evaluation method, providing two core metrics:

Similarity Calculation Rules

This benchmark uses a string similarity algorithm based on SequenceMatcher, with the following specific rules:

1. Empty Value Handling: One is empty and the other is not → similarity is 0
2. Complete Match: Both are identical (including both being empty) → similarity is 1
3. Case Insensitive: Convert to lowercase before comparison
4. Sequence Matching: Use longest common subsequence algorithm to calculate similarity (range: 0-1)

Similarity Score Interpretation:

• 1.0: Perfect match
• 0.8-0.99: Highly similar (may have minor formatting differences)
• 0.5-0.79: Partial match (extracted main information but incomplete)
• 0.0-0.49: Low similarity (extraction result differs significantly from ground truth)

1. Field-level Accuracy

Definition: The average similarity score for each metadata field.

Calculation Method:

Field-level Accuracy = Σ(similarity of that field across all samples) / total number of samples

Example: Suppose evaluating the title field on 100 samples, the sum of title similarity for each sample divided by 100 gives the accuracy for that field.

Use Cases:

• Identify which fields the model performs well or poorly on
• Optimize extraction capabilities for specific fields
• For example: If doi accuracy is 0.95 and abstract accuracy is 0.75, the model needs improvement in extracting abstracts

2. Overall Accuracy

Definition: The average of all evaluated field accuracies, reflecting the model's overall performance.

Calculation Method:

Overall Accuracy = Σ(field-level accuracies) / total number of fields

Example: Evaluating 7 fields (isbn, title, author, abstract, category, pub_time, publisher), sum these 7 field accuracies and divide by 7.

Use Cases:

• Provide a single quantitative metric for overall model performance
• Facilitate horizontal comparison between different models or methods
• Serve as an overall objective for model optimization

Using the Evaluation Script

compare.py provides a convenient evaluation interface:

from compare import main, write_similarity_data_to_excel

# Define file paths and fields to compare
file_llm = 'data/llm-label_textbook.jsonl'      # LLM extraction results
file_bench = 'data/benchmark_textbook.jsonl'     # Benchmark data

# For textbooks/ebooks
key_list = ['isbn', 'title', 'author', 'abstract', 'category', 'pub_time', 'publisher']

# For academic papers
# key_list = ['doi', 'title', 'author', 'keyword', 'abstract', 'pub_time']

# Run evaluation and get metrics
accuracy, key_accuracy, detail_data = main(file_llm, file_bench, key_list)

# Output results to Excel (optional)
write_similarity_data_to_excel(key_list, detail_data, "similarity_analysis.xlsx")

# View evaluation metrics
print("Field-level Accuracy:", key_accuracy)
print("Overall Accuracy:", accuracy)

Output Files

The script generates an Excel file containing detailed sample-by-sample analysis:

• sha256: File identifier
• origin_path: Original file path
• For each field (e.g., title):
  • llm_title: LLM extraction result
  • benchmark_title: Benchmark data
  • similarity_title: Similarity score (0-1)

📈 Statistics

Data Scale

First Batch (20250806):

• Ebooks: 70 records
• Academic Papers: 70 records
• Textbooks: 71 records
• Subtotal: 211 records

Second Batch (20251022):

• Ebooks: 354 records
• Academic Papers: 399 records
• Textbooks: 46 records
• Subtotal: 799 records

Total: 1010 benchmark test records

The data covers multiple languages (English, Chinese, German, Greek, etc.) and multiple disciplines, with both batches together providing a rich and diverse set of test samples.

🎯 Application Scenarios

1. LLM Performance Evaluation: Assess the ability of large language models to extract metadata from PDFs
2. Information Extraction System Testing: Test the accuracy of OCR, document parsing, and other systems
3. Model Fine-tuning: Use as training or fine-tuning data to improve model information extraction capabilities
4. Cross-lingual Capability Evaluation: Evaluate the model's ability to process multilingual literature

🔬 Data Characteristics

• ✅ Real Data: Real metadata extracted from actual PDF files
• ✅ Diversity: Covers literature from different eras, languages, and disciplines
• ✅ Challenging: Includes ancient texts, non-English literature, complex layouts, and other difficult cases
• ✅ Traceable: Each record includes SHA256 hash and original path

📋 Dependencies

pandas>=1.3.0
openpyxl>=3.0.0

Install dependencies:

pip install pandas openpyxl

🤝 Contributing

If you would like to:

• Report data errors
• Add new evaluation dimensions
• Expand the dataset

Please submit an Issue or Pull Request.

📧 Contact

If you have questions or suggestions, please contact us through Issues.

---

Last Updated: December 26, 2025