SpreadsheetLLM Compression Techniques

Implementation of compression techniques from the SpreadsheetLLM framework described in:

SpreadsheetLLM: Encoding Spreadsheets for Large Language Models
Yuzhang Tian, Jianbo Zhao, Haoyu Dong, et al.
Microsoft Research, arXiv:2407.09025 (2024)
📄 Paper Link

🎯 Implemented Techniques

1. Structural Anchors for Efficient Layout Understanding

• Identifies structural boundaries in spreadsheets
• Removes homogeneous regions that don't contribute to layout understanding
• Results: Up to 2.33x compression with 57.1% space saved

2. Inverted Index Translation

• Maps cell values to address ranges instead of row-by-row encoding
• Eliminates redundant encoding of repeated values
• Results: Up to 58.2% token reduction

3. Data Format Aggregation

• Groups cells by data format/type using rule-based recognition
• Recognizes 9 data types: Year, Integer, Float, Percentage, Scientific notation, Date, Time, Currency, Email, Others
• Results: Up to 80.4% space saved

📊 Performance Results

Technique	Original Tokens	Compressed	Compression Ratio	Space Saved
Structural Anchors	112	64	1.75x	42.9%
Inverted Index	55	23	2.39x	58.2%
Data Format Aggregation	51	10	5.10x	80.4%

🚀 Usage

Run the Jupyter Notebook:

jupyter notebook Implementation_of_Spreadsheet_compression_techniques.ipynb

The notebook contains detailed implementations, examples, and performance analysis for each compression technique.