nicolaaaaa/hemolab-bench-toy

Hemolab Bench — Toy

A small (1000 samples) subset of the full Hemolab Bench dataset, intended for quick-start examples and pipeline testing.

For the complete ~50 000-sample dataset see ibm-research/hemolab-bench.

How the toy was sampled

Samples were chosen as a uniform random subset of the full CSV id list:

import torch

all_ids = sorted(full_csv_ids)          # all ~50 000 ids, sorted
g = torch.Generator().manual_seed(1234) # independent seed for subsampling
perm = torch.randperm(len(all_ids), generator=g).tolist()
toy_ids = {all_ids[perm[i]] for i in range(500)}

Each selected id keeps its global train / validation / test label, which is assigned with the same rule as the full dataset (seed 42 — see below). This means the toy distribution across splits reflects the full dataset: roughly 80 % train, 10 % validation, 10 % test.

Because the subset is small, some alpha / reynolds bands may have very few (or zero) test rows. The toy is intended for code verification only.

Train / Validation / Test split

Both the toy and the full dataset share the same deterministic shuffle rule:

import torch

all_ids = sorted(full_csv_ids)
g = torch.Generator().manual_seed(42)
perm = torch.randperm(len(all_ids), generator=g).tolist()
shuffled = [all_ids[i] for i in perm]

n = len(shuffled)
train_ids      = shuffled[:int(n * 0.8)]
validation_ids = shuffled[int(n * 0.8):int(n * 0.9)]
test_ids       = shuffled[int(n * 0.9):]

The split is applied to the full CSV id list, so each id's label is stable whether you work with the full dataset or the toy.

Usage

from datasets import load_dataset

ds = load_dataset("nicolaaaaa/hemolab-bench-toy")
ds_train = ds["train"]
ds_test  = ds["test"]

sample = ds_test[0]
print(sample["alpha_split"], sample["reynold_split"])  # e.g. "ID", "IR"

Access a sample:

import numpy as np

points   = np.array(sample["points"]).reshape(25600, 3)
pressure = np.array(sample["pressure"])
wss      = np.array(sample["wss"]).reshape(25600, 3)

The shared mesh topology is stored in topology.parquet:

import pandas as pd

topology = pd.read_parquet(
    "hf://datasets/nicolaaaaa/hemolab-bench-toy/topology.parquet"
)
cells = topology["cells"].to_numpy().reshape(-1, 4)  # (9600, 4) quad indices

Dataset schema

Column	Type	Shape	Description
id	int32	—	Sample identifier
points	float32	76800 (= 25600×3, flattened)	Vertex coordinates
pressure	float32	25600	Per-vertex pressure
wss	float32	76800 (= 25600×3, flattened)	Per-vertex wall shear stress vector
alpha	float64	—	Geometry parameter
gammaY0..gammaY3	float64	—	Y-curvature parameters
gammaZ0..gammaZ3	float64	—	Z-curvature parameters
beta0..beta3	float64	—	Taper parameters
noise	float64	—	Geometry noise level
iseed	int64	—	Random seed used for geometry generation
reynolds	float64	—	Reynolds number
taylor	float64	—	Taylor number
alpha_split	string	—	Alpha band label (test rows only; null elsewhere)
reynold_split	string	—	Reynolds band label (test rows only; null elsewhere)

License

CDLA Permissive 2.0