rtemis demo

Synthetic tabular data for demonstrations and tests of supervised learning: 1,000 cases and 100 features of mixed type, of which 6 contribute to the outcome. The function of the features is known, so a fitted model can be compared with the truth as well as with held-out outcomes.

The dataset has two configs with the same features:

• regression: features x1 to x100 and the numeric outcome y.
• classification: features x1 to x100 and the binary outcome class.

The outcome is the last column of each table.

Structure

data/regression.parquet and data/classification.parquet: 1,000 rows, 101 columns each. Each config has one train split that holds all 1,000 cases.

The features form 10 blocks of 10 (x1 to x10, x11 to x20, and so on). Within each block:

Position	Type	Distribution
1 to 7	double	Standard normal; pairwise correlation 0.5 within the block
8	integer	Poisson; the mean differs by block (1 to 8)
9	categorical	3 to 5 equally likely levels, A, B, ...
10	binary categorical	Levels no and yes; yes with probability 0.2 to 0.5, by block

Features in different blocks are independent. In total: 70 numeric, 10 integer, 10 categorical, and 10 binary categorical features. There are no missing values. croissant.json gives the exact distribution of every feature.

Generating model

Six features contribute to the signal f:

f = 1.0 x3 + 0.8 x15 + 0.6 x3 x15
  + (0.5 + s[x29]) x24 + m[x29]
  + 0.25 (x38 - 3)
  + 0.7 [x50 = yes]

where [x50 = yes] is 1 when x50 is yes and 0 otherwise, and x29 has levels A, B, and C, with shifts m = (0, 0.5, -0.5) and slope changes s = (0, 1, -1).

Feature	Type	Role
x3	numeric	Main effect, and interaction with x15
x15	numeric	Main effect, and interaction with x3
x24	numeric	Slope 0.5, 1.5, or -0.5 by level of x29
x29	categorical	Shift by level, and modifies the slope of x24
x38	integer	Main effect
x50	binary categorical	Main effect

The other 94 features are noise. x3, x15, and x24 are each correlated 0.5 with six noise features of their block.

Outcomes:

• y = f + e, where e is normal with mean 0 and standard deviation 1. Population R-squared of f: 0.772.
• class is yes with probability logistic(1.5 (f - 0.35)), else no. Levels are no and yes; yes is the positive class. Bayes error: 0.181. Prevalence of yes: 0.456 in the data, 0.473 in the population.

Population quantities are estimated from 1,000,000 independent cases. The data were simulated in R with seed 2026.

Load

The Parquet files store each categorical column dictionary-encoded, with its levels in order, so Parquet readers return factors (R) or categorical columns (Python).

R - using rtemis:

dat <- rtemis::read("hf://datasets/rtemis/demo/data/classification.parquet")

R - using croissant, with categorical columns as factors in their declared level order:

library(croissant)
dat <- read_record_set("hf://datasets/rtemis/demo/croissant.json", "classification")

R - using nanoparquet:

url <- "https://huggingface.co/datasets/rtemis/demo/resolve/main/data/regression.parquet"
path <- file.path(tempdir(), "regression.parquet")
download.file(url, path, mode = "wb")
dat <- nanoparquet::read_parquet(path)

Python - using polars:

import polars as pl
dat = pl.read_parquet("hf://datasets/rtemis/demo/data/classification.parquet")

Metadata

croissant.json is the dataset's Croissant 1.1 metadata, written with the croissant R package. It describes each file with its SHA-256 checksum, each field with its distribution and its role in the generating model, and the Responsible AI properties of the data. Each categorical feature and the class outcome reference an enumeration record set (x9_levels, ..., x100_levels, class_levels) that lists their levels in order.

License

CC0 1.0: public domain dedication.