scale-align 0.1.0

Standard Classification Alignment & Local Enrichment - Automate audit and enrichment of industrial classifications

SCALE: Standard Classification Alignment & Local Enrichment

Automate the audit and enrichment of national classifications (e.g., KBLI) using international standards (e.g., ISIC, NACE), or other national standard classifications.

Features

• Bidirectional Asymmetric Retrieval: Captures hierarchical "entailment" relationships common in classifications
• Competitive Selection: Max-score + margin algorithm for controlled mapping while minimizing noise
• E5-Large-Instruct Model: Optimized for asymmetric retrieval tasks

Installation

pip install scale-align

Quick Start

scale-align \
    --source-dir /path/to/isic5 \
    --target-dir /path/to/nace21 \
    --correspondence correspondence.json \
    --output-dir ./output

Input Format

Classification Files

Text files with sentences separated by newlines:

/path/to/isic5/A.txt
/path/to/isic5/A0111.txt

Correspondence JSON

[["A","A"],["A0111","A01.11"],["A0112","A01.12"]]

Output Format

Each output file {source_code}_{target_code}.txt contains:

[0]:[0,1]:0.87439287
[1,2]:[2]:0.92
[3]:[]:0
[4]:[3]:0.9999999

• [source_indices]:[target_indices]:score
• Empty brackets [] indicate no alignment (gap detected)

CLI Options

Option	Default	Description
--threshold	0.7	Minimum similarity score
--margin	0.05	Margin from champion score
--batch-size	32	Embedding batch size
--device	auto	Device (cuda/cpu/auto)

License

MIT