README.md

metadata

language:
  - en
license: apache-2.0
task_categories:
  - tabular-classification
  - tabular-regression
tags:
  - tabular
  - synthetic
  - pretraining
  - in-context-learning
size_categories:
  - 100M<n<1B

Tabula Pretraining Corpus v2

A large-scale synthetic tabular dataset for pretraining transformer-based in-context learning models for tabular data (similar to TabPFN).

Overview

Metric	Value
Total rows	272,271,776
Total datasets	10,867
Shards	135
Mean utility AUC	0.851
Format	Parquet (float32)

Schema

Each shard is a Parquet file with a fixed-width schema:

• feat_0 through feat_63: Float32 feature columns. Unused slots are NaN.
• target: Float32 target variable (classification label or regression target).
• _source_meta: JSON string with dataset metadata including:
  • generator: Which synthetic generator produced this dataset
  • task_type: "binary", "multiclass", or "regression"
  • n_features: Number of active features (rest are NaN-padded)
  • n_classes: Number of target classes
  • n_samples: Number of rows in the original dataset
  • domain: Semantic domain (finance, health, etc.)
  • feature_names: Original domain-specific column names

Generators

Generator	Datasets
GaussianMixture	3,029
Polynomial	2,738
SCM	2,674
TreePrior	2,096
Regression	325
MixedType_GaussianMixture	2
MixedType_SCM	2
MixedType_TreePrior	1

Task Types

Type	Datasets
binary	8,396
multiclass	2,146
regression	325

Domains

Domain	Datasets
hr	1,033
education	1,031
telecom	1,028
science	1,020
iot	1,005
finance	1,000
health	985
ecommerce	977
logistics	972
environment	935
manufacturing	881

Quality Gates

Every generated dataset passes quality gates before inclusion:

• No constant columns — all features must vary
• No all-null columns
• Minority class fraction ≥ 5% for classification
• Duplicate row fraction ≤ 30%
• RF utility AUC ≥ 0.55 — a Random Forest must achieve above-chance cross-validated AUC

Gate failure rate: 22.4%

Data Augmentation

• Missingness injection: ~30% of datasets have random missing values injected
• Concept drift: ~20% of datasets have feature distribution shifts

Usage

from datasets import load_dataset

ds = load_dataset("avewright/tabula-pretraining-corpus-v2", split="train", streaming=True)

for batch in ds.iter(batch_size=512):
    features = batch["feat_0"]  # access individual features
    target = batch["target"]
    meta = batch["_source_meta"]  # JSON metadata string

License

Apache 2.0