README.md

metadata

pretty_name: SALT-KG
license: cc-by-nc-sa-4.0
knowledge_graph:
  path: data/salt-kg/salt-kg.json
  format: json
  description: >-
    Canonical full Operational Business Knowledge Graph metadata file to use
    programmatically.
viewer_derivatives:
  - path: data/salt-kg/salt-kg-summary.parquet
    purpose: >-
      Generated Hugging Face Dataset Viewer summary file only. Not the canonical
      metadata source.
  - path: data/salt-kg/salt-kg-fields.parquet
    purpose: >-
      Generated Hugging Face Dataset Viewer field-level display file only. Not
      the canonical metadata source.
configs:
  - config_name: salt_kg_summary
    default: true
    data_files:
      - split: metadata
        path: data/salt-kg/salt-kg-summary.parquet
  - config_name: salt_kg_fields
    data_files:
      - split: metadata
        path: data/salt-kg/salt-kg-fields.parquet
  - config_name: sales_document
    data_files:
      - split: full
        path: data/salt/I_SalesDocument.parquet
      - split: train
        path: data/salt/I_SalesDocument_train.parquet
      - split: test
        path: data/salt/I_SalesDocument_test.parquet
  - config_name: sales_document_item
    data_files:
      - split: full
        path: data/salt/I_SalesDocumentItem.parquet
      - split: train
        path: data/salt/I_SalesDocumentItem_train.parquet
      - split: test
        path: data/salt/I_SalesDocumentItem_test.parquet
  - config_name: joined_tables
    data_files:
      - split: train
        path: data/salt/JoinedTables_train.parquet
      - split: test
        path: data/salt/JoinedTables_test.parquet
  - config_name: customer
    data_files:
      - split: full
        path: data/salt/I_Customer.parquet
  - config_name: address_org_name_postal_address
    data_files:
      - split: full
        path: data/salt/I_AddrOrgNamePostalAddress.parquet

SALT-KG: A Benchmark for Semantics-Aware Learning on Enterprise Tables

Description

This repository contains the dataset from our paper SALT-KG: A Benchmark for Semantics-Aware Learning on Enterprise Tables, presented at EurIPS'25 Table Representation Workshop.

The canonical metadata file for SALT-KG is data/salt-kg/salt-kg.json. If you want to use the full Operational Business Knowledge Graph metadata, use this file directly.

The files data/salt-kg/salt-kg-summary.parquet and data/salt-kg/salt-kg-fields.parquet are generated helper files for the Hugging Face Dataset Viewer only. They are provided to make the metadata browsable on Hugging Face and are not replacements for data/salt-kg/salt-kg.json.

The parquet files under data/salt/ contain the relational SALT tables that remain available in the Dataset Viewer as separate table configurations.

Abstract

Building upon the SALT benchmark for relational prediction, we introduce SALT-KG, a benchmark for semantics-aware learning on enterprise tables. SALT-KG extends SALT by linking its multi-table transactional data with a structured Operational Business Knowledge represented in a Metadata Knowledge Graph (OBKG) that captures field-level descriptions, relational dependencies, and business object-types. This extension enables evaluation of models that jointly reason over tabular evidence and contextual semantics—an increasingly critical capability for foundation models on structured data. Empirical analysis reveals that while metadata-derived features yield modest improvements in classical prediction metrics, these metadata features consistently highlight gaps in models’ ability to leverage semantics in relational context. By reframing tabular prediction as semantics-conditioned reasoning, SALT-KG establishes a benchmark to advance tabular FMs grounded in declarative knowledge, providing the first empirical step toward semantically linked tables in structured data at enterprise scale.

Why SALT-KG

There is growing research on Tabular Foundation Models. TRL models are typically trained and evaluated on benchmarks that represent relational structure but lack explicit semantic grounding or declarative context. Knowledge graph (KG) and data integration communities have explored connecting tables to semantic graphs through systems such as JENTAB. We can bridge this gap by enriching enterprise relational data with an explicit semantic layer that links tables, fields, and business objects through declarative knowledge in KG

How was SALT-KG Created

For every relation (Table) in the underlying SALT dataset, we find a matching node in the KG (a View).. We extract triples related to the Views that include: Fields: data abstraction nodes with associated fields, labels, associations, data classes, reference fields, and other elements. ObjectNodeTypes: further semantic metadata through technical definitions, business object descriptions

Dataset Overview

The SALT-KG dataset consists of 4 tables from the SALT benchmark, enriched with semantic metadata from an Operational Business Knowledge Graph (OBKG). The dataset includes:

• 4 relational tables with transactional data
• Metadata Knowledge Graph (OBKG) with field-level descriptions, relational dependencies, and business object types
• Canonical metadata JSON at data/salt-kg/salt-kg.json
• Generated Hugging Face viewer helper files at data/salt-kg/salt-kg-summary.parquet and data/salt-kg/salt-kg-fields.parquet
• Train and test splits for the benchmark tables where provided, plus full-table parquet files for direct browsing

Requirements

N/A

Known Issues

No known issues

Authors

• Isaiah Onando Mulang'
• Felix Sasaki
• Tassilo Klein
• Jonas Kolk
• Nikolay Grechanov
• Johannes Hoffart

Citations

If you use this dataset in your research, please cite the following paper:

@inproceedings{mulang2025saltkg,
  title={SALT-KG: A Benchmark for Semantics-Aware Learning on Enterprise Tables},
  author={Mulang', Isaiah Onando and Sasaki, Felix and Klein, Tassilo and Kolk, Jonas and Grechanov, Nikolay and Hoffart, Johannes},
  booktitle={Proceedings of the AI for Tabular Data workshop at EurIPS 2025},
  year={2025}
}

How to obtain support

Create an issue in this repository if you find a bug or have questions about the content.

For additional support, ask a question in SAP Community.

Contributing

If you wish to contribute code, offer fixes or improvements, please send a pull request. Due to legal reasons, contributors will be asked to accept a DCO when they create the first pull request to this project. This happens in an automated fashion during the submission process. SAP uses the standard DCO text of the Linux Foundation.

License

Copyright (c) 2025 SAP SE or an SAP affiliate company. All rights reserved. This project is licensed under the CC-BY-NC-SA-4.0 except as noted otherwise in the LICENSE file.

SAP expressly reserves its rights against text and data mining for commercial purposes as described in TDM_RESERVATION.txt.