STRABLE: Benchmarking Tabular Machine Learning with Strings

STRABLE is a benchmarking corpus of 108 real-world tabular datasets containing string features, designed to support empirical research on tabular machine learning pipelines that handle string entries.

Dataset Summary

Existing tabular benchmarks either exclude string columns or flatten them into fixed numerical representations before evaluation, preventing the study of alternative string-handling strategies. STRABLE fills this gap with raw, minimally preprocessed tables spanning 8 application fields, covering 13 binary classification, 19 multi-class classification, and 76 regression tasks.

Key Statistics

• 108 datasets from 33 distinct public sources
• 8 application fields: Economy, Health, Infrastructure, Energy, Education, Commerce, Food, Social
• Median: 7,700 rows · 18 columns · 17-character strings per cell · 1,200 unique values per string column
• String taxonomy: Categorical (50%), Names (23%), Structured Codes (17%), Free Text (8%), Identifiers (0.5%)

Loading a Dataset

Each of the 108 datasets is stored as a parquet file in its own subdirectory.

import pandas as pd

# Load one dataset directly
df = pd.read_parquet(
    "hf://datasets/strablebench/STRABLE/beer-ratings/data.parquet"
)

# Or via the HuggingFace datasets library
from datasets import load_dataset
ds = load_dataset("strablebench/STRABLE", data_dir="beer-ratings")
df = ds["train"].to_pandas()

Available dataset names (subdirectory keys) are listed in the repository file tree, and described in full in Appendix C.4 of the paper.

String Taxonomy

Type	Description	Example
Categorical	Low-uniqueness repeating labels	"General Acute Care", "Red"
Name	Proper nouns for entities	"John Doe", "Max Mara"
Structured Code	Strings with recognizable patterns	ZIP codes, ICD codes, dates
Free Text	Multi-word natural language prose	User reviews, clinical notes
Identifier	Near-unique opaque keys	UUIDs, auto-generated IDs

Preprocessing

STRABLE applies minimal preprocessing to reflect data as found in practice:

• Flattening of nested structures; removal of duplicate rows, single-value columns, all-null columns
• No feature engineering; no missing values imputation
• Regression targets skewness-minimized via a candidate set of transformations
• Large tables subsampled to 75,000 rows (stratified for classification)

Citation

@inproceedings{strable2026neurips,
  title     = {{STRABLE}: Benchmarking Tabular Machine Learning with Strings},
  booktitle = {Anonimous Conference},
  year      = {2026},
}

License

Individual source datasets retain their original licenses; see Appendix C.4 of the paper for per-dataset attribution and source URLs.