trl-bench-ckpts

TRL-Bench pretrained checkpoints

Pretrained weights for models evaluated in TRL-Bench (paper: arXiv:2606.09323; code: LOGO-CUHKSZ/TRL-Bench). Each subdirectory contains:

• a copy of the upstream LICENSE
• a NOTICE file with citation, provenance, and SHA256
• the binary checkpoint(s)

Most subdirectories mirror an upstream release. TabSketchFM is the exception: its authors published no checkpoint, so the weights here were trained by the TRL-Bench authors — see below.

The mirrored upstream checkpoints (TUTA, TURL, TaBERT) have not been modified — they are byte-identical to the upstream releases. Those mirrors exist only to provide a reliable, scriptable download path that does not depend on third-party file-share services (Google Drive, SharePoint) which can rate-limit or expire links.

Coverage

TaBERT is mirrored under the upstream CC BY-NC 4.0 terms strictly for non-commercial reproduction of the paper experiments. If you intend a commercial application you may not use that checkpoint; train your own or seek an alternate license from the original authors (see tabert/NOTICE).

TabSketchFM is not a mirror. IBM released the TabSketchFM code but no pretrained weights and no pretraining corpus. The checkpoint here was trained from scratch by the TRL-Bench authors using the upstream pretrain.py, and is the checkpoint behind the TabSketchFM results in the TRL-Bench paper. It is published because no upstream checkpoint exists and our results are otherwise irreproducible. Do not attribute its behaviour, quality, or results to IBM. No TabSketchFM source code is redistributed here — obtain the code from IBM/tabsketchfm under the upstream CC BY-NC-ND 4.0 license. Full provenance is in tabsketchfm/NOTICE.

Other TRL-Bench models — Starmie (user-trained) and TABBIE (obtain from upstream) — are not distributed here. See docs/CHECKPOINT_LICENSES.md and scripts/download_checkpoints.sh in the TRL-Bench repo for the upstream paths used for those models.

Downloading

Programmatic (huggingface_hub):

from huggingface_hub import snapshot_download
snapshot_download(
    "logo-lab/trl-bench-ckpts",
    allow_patterns=["tuta/*", "turl/*", "tabert/*", "tabsketchfm/*"],
    local_dir="./checkpoints",
)

Or via TRL-Bench's bundled script (verifies SHA256 against scripts/checksums.sha256 after download):

git clone https://github.com/LOGO-CUHKSZ/TRL-Bench.git
cd TRL-Bench
bash scripts/download_checkpoints.sh

Layout

tuta/
  LICENSE              # MIT (upstream Microsoft)
  NOTICE               # citation, provenance, SHA256
  tuta.bin             # 511 MB
turl/
  LICENSE              # Apache-2.0 (upstream sunlab-osu)
  NOTICE               # citation, provenance, SHA256
  pretrained/
    config.json
    pytorch_model.bin  # 1.2 GB
tabert/
  LICENSE              # CC BY-NC 4.0 (upstream Facebook AI Research)
  NOTICE               # citation, provenance, SHA256 + NC terms
  tabert_base_k3/
    model.bin          # 266 MB
    tb_config.json
    version.txt
tabsketchfm/
  LICENSE              # CC BY-NC-ND 4.0 (upstream IBM; covers the code)
  NOTICE               # provenance: TRL-Bench-trained, NOT an IBM release
  epoch=10-step=27786.ckpt   # 1.3 GB

These paths match the expectations of TRL-Bench's scripts/checksums.sha256 and the per-model wrappers in src/trl_bench/models/{tuta,turl,tabert,tabsketchfm}/.

Attribution

For TUTA, TURL, and TaBERT this is a third-party mirror: the upstream authors are the sole copyright holders of those model weights.

For TabSketchFM the weights were produced by the TRL-Bench authors; IBM holds the architecture, training code, and method, and is credited accordingly. IBM did not produce, review, or endorse this checkpoint.

The per-model LICENSE and NOTICE files in each subdirectory are authoritative.

If you use these checkpoints, cite the upstream papers (see each subdir's NOTICE). If you use TRL-Bench, also cite:

@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}
}