Maiernator/ctboost-tabarena-lite-0.1.56

CTBoost 0.1.56 — TabArena-v0.1 Lite artifacts

This repository contains the auditable artifacts for a local TabArena-v0.1 Lite evaluation of ctboost==0.1.56, produced for autogluon/tabarena PR #479.

This is a benchmark-artifact repository, not a training dataset.

Result

Metric	Value
Lite Elo	1166.7 (+52.1 / -67.5)
Win rate	0.3957
Imputed CTBoost tasks	0 / 51
Position among config-default rows	23 / 38

Default-to-default reference points from the same evaluator run were CatBoost 1337.5, XGBoost 1183.9, CTBoost 1166.7, and LightGBM 1153.3.

Elo is roster-dependent. These numbers use the 42-method TabArena artifact collection present at commit 31026f7d758390994353eba79fbfa6747616f365 (85 evaluated rows including default, tuned, tuned ensemble, and system entries).

Protocol and scope

• TabArena-v0.1 Lite: all 51 official datasets, split r0f0 only.
• Problem types: 30 binary classification, 8 multiclass classification, and 13 regression tasks.
• One measured configuration: CTBoost_c1_default_BAG_L1.
• Default-only: no HPO configurations or TabArena-Full results are included or claimed.
• Eight fitted bag children per task: 51 parents and 408 child fits.
• Released CTBoost version: 0.1.56.
• TabArena integration branch: ctboost-integration-0156.
• TabArena/PR commit used for fitting and evaluation: 31026f7d758390994353eba79fbfa6747616f365.
• AutoGluon Tabular: 1.6.2b20260821; TabArena: 0.0.1; Python: 3.12.13.

The run-level metadata.yaml correctly records can_hpo: false because this artifact contains only the measured default configuration. The integration in PR #479 separately supports a frozen 200-configuration HPO portfolio; that portfolio was not executed here. Consequently, hpo_results.parquet contains the default result row only despite its canonical TabArena filename.

Validation

• Six of six final Kaggle shards completed with exit code 0 and no fatal error.
• Exact 51-task official Lite coverage: no missing, extra, duplicate, or path-colliding results.
• All 51 raw results.pkl files and all six source shard archives matched their manifest sizes and SHA-256 hashes.
• All prediction arrays and validation/test metrics were finite and shape-consistent.
• Classification probability outputs were valid.
• TabArena raw processing inferred one complete default row over 51 datasets with zero imputation.
• Direct context scoring and the native raw-processing/evaluation path independently produced the same 1166.7 Elo result.

Hardware disclosure

The six shards ran on Kaggle CPU instances requesting 4 CPUs, 28 GB RAM, and a 3,600-second limit. The canonical TabArena disclosure is 8 CPUs, 32 GB RAM, and 3,600 seconds. Accuracy/Elo is provided for review, but the recorded Kaggle timing must not be treated as an official comparable TabArena runtime result.

Files

canonical-results/
  metadata.yaml
  results/hpo_results.parquet
  results/model_results.parquet
reports/
  ctboost_lite_summary.json
  leaderboard_lite.csv
validation/
  aggregate_manifest.json
  manifests/shard-0.json ... shard-5.json
  environment/shard-0-pip-freeze.txt ... shard-5-pip-freeze.txt
SHA256SUMS

Load the canonical result tier after downloading the repository:

from pathlib import Path

from huggingface_hub import snapshot_download
from tabarena.models._method_metadata import MethodMetadata

root = Path(
    snapshot_download(
        repo_id="Maiernator/ctboost-tabarena-lite-0.1.56",
        repo_type="dataset",
    )
)
method = MethodMetadata.from_yaml(path=root / "canonical-results" / "metadata.yaml")
results = method.load_results()
print(results)

Verify the publication bundle with sha256sum -c SHA256SUMS (or the platform-equivalent SHA-256 tool).

Citation and provenance

The benchmark protocol, evaluator, and reference artifacts come from TabArena (paper). Dataset curation and source provenance are documented by the official TabArena dataset-curation project. The exact integration and evaluation revision is 31026f7d758390994353eba79fbfa6747616f365, associated with autogluon/tabarena PR #479.

If you use these artifacts in a publication, cite Nick Erickson, Lennart Purucker, Andrej Tschalzev, David Holzmüller, Prateek Mutalik Desai, David Salinas, and Frank Hutter, “TabArena: A Living Benchmark for Machine Learning on Tabular Data,” Advances in Neural Information Processing Systems 38 (2026), arXiv:2506.16791.

Licensing and raw-data notice

CTBoost and TabArena source code are Apache-2.0 licensed. The benchmark tasks originate from 51 third-party OpenML datasets, each of which retains its original authorship and license. The public bundle therefore uses Hugging Face's other license designation and contains only aggregate result tables, run metadata, manifests, environment records, and checksums.

The raw results.pkl archive is intentionally not published here because it contains target arrays derived from the underlying datasets. It can be transferred to TabArena maintainers through an approved artifact-ingestion channel if they request it and confirm the appropriate handling of the source-dataset licenses.