TRL-Rbench

Paper: arXiv:2606.09323 — TRL-Bench: Standardizing Cross-Paradigm Representation-Level Evaluation of Tabular Encoders · Code: LOGO-CUHKSZ/TRL-Bench

Row-level evaluation suite of TRL-Bench:

• row_prediction: 50 OpenML tables, 1.1M rows (887,720 train / 110,962 validation / 110,983 test) across 123 hand-verified targets. Per-row schema bundles features and targets as JSON dicts so all 50 tables share one config.
• record_linkage: 16 entity-matching sources (8 clean DeepMatcher + 4 dirty DeepMatcher + 4 WDC LSPM v2 sizes) unified into one config with 357,833 train / 95,817 validation / 42,835 test row pairs. Per-row stores both records as JSON dicts plus source / family / pair_id / label.

Schemas

row_prediction

{
  "openml_id": int32,                 # e.g., 3, 38, 458, ..., 46980 (50 distinct IDs)
  "dataset_name": string,             # e.g., "kc2"
  "row_idx": int32,                   # row index within the original OpenML CSV
  "record_json": string,              # JSON dict of feature columns
  "targets_json": string,             # JSON dict of target columns
  "target_specs_json": string,        # JSON list of {name, role, task_type}
  "dataset_metadata_json": string,    # full dataset.json copy (label_columns, splits, fingerprint)
}

To filter to one OpenML table:

from datasets import load_dataset
import json
ds = load_dataset("logo-lab/trl-rbench", "row_prediction")
ds_3 = ds.filter(lambda x: x["openml_id"] == 3)
print(json.loads(ds_3["train"][0]["record_json"]))

record_linkage

{
  "source": string,                   # e.g., "deepmatcher_beer", "wdc_products_xlarge"
  "family": string,                   # "deepmatcher_clean" | "deepmatcher_dirty" | "wdc_products"
  "pair_id": string,                  # f"{source}/{split}/{idx}"
  "table_a_record_json": string,      # JSON dict of left record
  "table_b_record_json": string,      # JSON dict of right record
  "label": int32,                     # 0 = non-match, 1 = match
}

Label-equivalent columns documented in the paper appendix (cluster_id, identifiers on WDC; class on Fodors-Zagats) are NOT removed at the parquet layer; users who want the leakage-free view should drop these keys after json.loads(record_json).

Quickstart

from datasets import load_dataset
import json

# Row prediction
rp = load_dataset("logo-lab/trl-rbench", "row_prediction")
print(rp)
sample = rp["train"][0]
print(json.loads(sample["record_json"]))
print(json.loads(sample["targets_json"]))
print(json.loads(sample["target_specs_json"]))

# Record linkage
rl = load_dataset("logo-lab/trl-rbench", "record_linkage")
sample = rl["train"][0]
print(sample["source"], sample["label"])
print(json.loads(sample["table_a_record_json"]))
print(json.loads(sample["table_b_record_json"]))

# Filter to one source
rl_beer = rl.filter(lambda x: x["source"] == "deepmatcher_beer")

See LICENSES.md for per-source license details.

Citation

If you use this dataset, please cite the TRL-Bench paper:

@article{pang2026trl,
  title={TRL-Bench: Standardizing Cross-Paradigm Representation-Level Evaluation of Tabular Encoders},
  author={Pang, Wei and Jian, Xiangru and Li, Hehan and Yu, Zhixuan and Xue, Alex and Li, Jinyang and Dong, Zhengyuan and Zhao, Xinjian and Xu, Hao and Zhang, Chao and Cheng, Reynold and {\"O}zsu, M. Tamer and Yu, Tianshu},
  journal={arXiv preprint arXiv:2606.09323},
  year={2026}
}