TRL-Rbench: A Cross-Paradigm Benchmark for Evaluating Tabular Encoders

TRL-Rbench is a standardized evaluation suite for tabular representation learning models. The dataset includes a row_prediction component with 50 OpenML tables and 1.1 million rows across 123 hand-verified targets. It was created by logo-lab and last updated on June 11, 2026.

Use Cases

• Benchmarking tabular encoder performance based on the row_prediction task across 50 tables.
• Evaluating representation quality for record linkage tasks based on the 16 entity-matching sources.
• Standardizing cross-paradigm evaluation of machine learning models based on the described suite structure.

Strengths

• Includes 50 OpenML tables for row_prediction, providing a diverse testbed.
• Contains 1.1 million total rows (887,720 train / 110,962 validation / 110,983 test) for model training and evaluation.
• Features 123 hand-verified targets, suggesting a level of quality control.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the record_linkage component is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

logo-lab on Hugging Face, associated with the paper 'TRL-Bench: Standardizing Cross-Paradigm Representation-Level Evaluation of Tabular Encoders'.

Collection Method

Likely compiled from 50 OpenML datasets and 16 entity-matching sources for benchmarking purposes.

Freshness

Last updated 2026-06-11 03:52:46; freshness should be verified.