Datasetpublic

CSE472-blanket-challenge/SCM3K

SCM3K Benchmark dataset for the paper: The Good, the Bad, and the Ugly of Markov Boundary for Tabular Prediction Shu Wan, Abhinav Gorantla, Huan Liu, K. Selçuk Candan 3,450 tabular prediction tasks sampled from random structural causal models (SCMs), totalling 3.45M records (1,000 samples per task). Each task ships with the ground-truth Markov boundary of the target node, so you can evaluate feature selection and prediction under known causal structure. Nine feature-count… See the full description on the dataset page: https://huggingface.co/datasets/CSE472-blanket-challenge/SCM3K.

sourceHugging Facemitupdated 4mo agoView on Hugging Face

0likes35downloads

Dataset Card

SCM3K

Benchmark dataset for the paper:

[The Good, the Bad, and the Ugly of Markov Boundary for Tabular Prediction](https://huggingface.co/papers/2605.29411) Shu Wan, Abhinav Gorantla, Huan Liu, K. Selçuk Candan

3,450 tabular prediction tasks sampled from random structural causal models (SCMs), totalling 3.45M records (1,000 samples per task). Each task ships with the ground-truth Markov boundary of the target node, so you can evaluate feature selection and prediction under known causal structure. Nine feature-count levels from 40 to 1,000.

Splits

One HF split per feature count F. No predefined train/test partition — use HF slice syntax (e.g. split="f200[:80%]").

Split	`F`	`num_nodes`	DAG density	MB-ratio band	Tasks
f40	40	41	ER `[0.2, 0.4]`	`[0.10, 0.90]`	300
f60	60	61	ER `[0.2, 0.4]`	`[0.10, 0.90]`	300
f80	80	81	ER `[0.2, 0.4]`	`[0.10, 0.90]`	300
f100	100	101	ER `[0.2, 0.4]`	`[0.10, 0.90]`	300
f200	200	201	ER `[0.01, 0.02, 0.04]`	`[0.05, 0.95]`	450
f400	400	401	ER `[0.01, 0.02, 0.04]`	`[0.05, 0.95]`	450
f600	600	601	ER `[0.01, 0.02, 0.04]`	`[0.05, 0.95]`	450
f800	800	801	ER `[0.01, 0.02, 0.04]`	`[0.05, 0.95]`	450
f1000	1000	1001	ER `[0.01, 0.02, 0.04]`	`[0.05, 0.95]`	450
				Total	3,450

Row schema

Each row is one prediction task.

Field	Type	What it stores
`task_id`	string	unique task identifier
`X`	list\<list\<f32\>\>	feature matrix, 1,000 x F
`y`	list\<f32\>	target vector, length 1,000
`n_features`	int	number of features (= F)
`mb_mask`	list\<bool\>	true Markov boundary mask over features
`mb_ratio`	float	fraction of features in the Markov boundary
`density`	float	edge density of the generating DAG
`dag`	struct	`dag_id`, `num_nodes`, `edge_list`
`scm`	struct	`scm_type`, `noise_model`, `pnl`, `coeff_range`, `noise_std`
`target`	struct	`target_node`, `parent_mask`, `child_mask`, `spouse_mask`
`meta`	struct	`n_samples`, `seed`

How the data was generated

DAGs: Erdos-Renyi, 5 graphs per (num_nodes, density) pair, seed 42.

SCMs: six families — LINEAR_GAUSSIAN, LINEAR_NONGAUSSIAN, NL_ANM_GAUSSIAN, NL_ANM_NONGAUSSIAN, PNL, HETEROSKEDASTIC. Each DAG gets 5 SCM instantiations with n_samples=1000, coeff_range=1.0, noise_std=0.5.

Quick start

python

from datasets import load_dataset

ds = load_dataset("CSE472-blanket-challenge/SCM3K", split="f200")
task = ds[0]

X  = task["X"]          # 1000 x 200
y  = task["y"]          # 1000
mb = task["mb_mask"]    # ground-truth Markov boundary

Citation

bibtex

@article{wan2026gbu,
  title  = {The Good, the Bad, and the Ugly of Markov Boundary
            for Tabular Prediction},
  author = {Wan, Shu and Gorantla, Abhinav and Liu, Huan
            and Candan, K. Sel{\c{c}}uk},
  year   = {2026},
}