SemiHS-Bench: A Semiconductor Benchmark for HS6 Classification from Sparse ERP-Style Text

SemiHS-Bench is a semiconductor benchmark for HS6 customs classification from sparse ERP-style text. This Version 2.0.0 data release contains 1,800 benchmark records: 900 training records and 900 evaluation records, covering HS2022 classifications across 38 HS4 families and 73 HS6 codes. It also includes an 889-entry reference-ruling corpus from CROSS, EBTI, and Japan Customs, together with HS6 descriptions, taxonomy files, JSON Schemas, a SHA-256 distribution manifest, statistics, and documentation.

The benchmark records are compiled from manufacturer-catalog and bill-of-lading line-item sources and were expert-validated as described in the accompanying datasheet and balance report. Raw source dumps are not redistributed. The source repository and LICENSE-DATA file describe the upstream provenance and attribution requirements.

Intended use includes research on customs classification, retrieval-augmented classification, benchmark evaluation, and analysis of sparse product descriptions. The dataset is not a substitute for official customs advice. The source materials note that some source_reference and justification_text fields may retain supplier part numbers, and reference-ruling excerpts may retain names or contact details appearing in public rulings. Users should review the accompanying documentation and comply with applicable privacy, rights, and attribution requirements.

Source repository: https://github.com/cti-iel/SemiHS-Bench
Zenodo DOI: https://doi.org/10.5281/zenodo.21891036

Associated article: This dataset supports an article selected for presentation at CIKM 2026.