LeaderboardVersion history
STRABLE was accepted to the NeurIPS 2026 Evaluation and Datasets track. The board picks up from the paper's Pareto plot (75 of 445 pipelines, Figure 4a) and grows as new models come out.
Median over 108 tables × 3 folds (R² for regression, AUC for classification) · run time per 1K samples, log scale
| Rank | Pipelinedefault configs unless tuned | Scoremedian | Timetotal s / 1K samples, median |
|---|
Could not load the results. Try ↻ Refresh, or reload the page later.
Download the current board, with the version and date, to include it in a paper. Current view:
The plot follows the Task, Score, Encoders and Compute chosen above; the CSV, JSON and XLSX rankings and the learner × encoder matrix (one row per learner, one column per encoder, - where the view has no such pipeline, end-to-end models left out) follow them and the Sort. File names carry these choices. The Parquet file holds all three views at once, one row per pipeline and view. The raw results hold every fold: one row per pipeline, table and outer fold, with its score and run time.
What changed in each version. To cite a rank:
Help us improve STRABLE
Tell us in a GitHub issue: a feature, a bug, or a question about the board.
Open an issue@misc{blayer2026strablebenchmarkingtabularmachine,
title={STRABLE: Benchmarking Tabular Machine Learning with Strings},
author={Gioia Blayer and Myung Jun Kim and Félix Lefebvre and Lennart Purucker and Alan Arazi and Eilam Shapira and Roi Reichart and Frank Hutter and Marine Le Morvan and David Holzmüller and Gaël Varoquaux},
year={2026},
eprint={2605.12292},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2605.12292},
}