quantem-data

Reference electron-microscopy datasets for browsing and learning. Open in your browser via quantem.widget — no quantem.live install needed.

Two buckets:

• 4dstem/ — 4D-STEM acquisitions. _npy_bin* variants are pre-binned NumPy files for fast workshop / Colab demos; the originals are full Arina h5 bundles.
• haadf/ — HAADF survey images. _npy variants are pre-cooked NumPy + a meta.json sidecar carrying sampling + optics; the originals are full Velox EMD files.

A ready-to-run notebook sits at notebooks/show4dstem_colab.ipynb in this repo.

One-click workshop notebook (Google Colab)

It does the full pipeline in 7 cells: install quantem.widget (TestPyPI rc) + quantem (dev fork branch), download the pre-binned NumPy bundle from this dataset, wrap it as Dataset4dstem.from_tensor, render with Show4DSTEM in your browser via WebGPU. No CUDA on Colab. No quantem.live.

Workshop v1 — real gold 4D-STEM: browse + bright field + dark field + probe + DPC, all on the Colab T4. No quantem.live, no local install. The notebook lives at notebooks/berk_workshop_v1.ipynb in this repo and on the berk-workshop branch of bobleesj/quantem.

Workshop quick start — Show4DSTEM (any Jupyter)

!pip install -q --pre --extra-index-url https://test.pypi.org/simple/ quantem.widget huggingface_hub
!pip install -q git+https://github.com/electronmicroscopy/quantem.git@dev

import os, json, numpy as np, torch
from huggingface_hub import snapshot_download
from quantem.core.datastructures import Dataset4dstem
from quantem.widget import Show4DSTEM

folder = snapshot_download("bobleesj/quantem-data", repo_type="dataset",
                          allow_patterns=["4dstem/gold_512_npy_bin8/*"])
asset = os.path.join(folder, "4dstem", "gold_512_npy_bin8")
data = np.load(os.path.join(asset, "data.npy"))
meta = json.load(open(os.path.join(asset, "meta.json")))

dset = Dataset4dstem.from_tensor(torch.from_numpy(data),
                                 sampling=meta["sampling"], units=meta["units"])
Show4DSTEM(dset)

Workshop quick start — Show2D for HAADF

import os, json, numpy as np, torch
from huggingface_hub import snapshot_download
from quantem.widget import Show2D

folder = snapshot_download("bobleesj/quantem-data", repo_type="dataset",
                          allow_patterns=["haadf/gold_haadf_npy/*"])
asset = os.path.join(folder, "haadf", "gold_haadf_npy")
img  = np.load(os.path.join(asset, "data.npy"))
meta = json.load(open(os.path.join(asset, "meta.json")))

Show2D(torch.from_numpy(img), sampling=meta["sampling"], units=meta["units"])

Acquisition parameters

The 20260423 drift session's optics are confirmed via the session's own HAADF EMD (AccelerationVoltage, BeamConvergence, CameraLength). 4D-STEM scan_sampling is an operator pattern from a sibling SSB session — the drift acquisition itself was never per-file calibrated.

dataset	voltage	probe	CL	scan	scan sampling	det pitch	mag
haadf/gold_haadf_npy	300 kV ✓	30 mrad ✓	91 mm ✓	4096² image	0.0186 nm/px	n/a	FOV 76.2 nm
4dstem/gold_512_npy_bin8	300 kV ✓	30 mrad ✓	91 mm ✓	512²	0.5 Å (op)	3.68 mrad/px	unknown
4dstem/gold_512_npy_bin4	300 kV ✓	30 mrad ✓	91 mm ✓	512²	0.5 Å (op)	1.84 mrad/px	unknown
4dstem/gold_512	300 kV ✓	30 mrad ✓	91 mm ✓	512²	0.5 Å (op)	0.46 mrad/px	unknown
4dstem/gold_30mrad1.3mx04…09	300 kV	30 mrad	91 mm	smaller	(session yaml)	0.46 mrad/px	1.3 Mx
haadf/gold_haadf.emd	300 kV ✓	30 mrad ✓	91 mm ✓	4096² image	0.0186 nm/px	n/a	FOV 76.2 nm

✓ = confirmed via EMD/yaml. (op) = operator pattern, not file-certified.

Datasets at a glance

name	kind	shape	dtype	size	use
4dstem/gold_512_npy_bin8/	NumPy bundle	(512, 512, 24, 24)	uint16	~302 MB	workshop / Colab demo
4dstem/gold_512_npy_bin4/	NumPy bundle	(512, 512, 48, 48)	uint16	~1.2 GB	sharper workshop version
4dstem/gold_512/	Arina h5	(512, 512, 192, 192)	uint16	~5 GB	power user
4dstem/gold_30mrad1.3mx04 … 09	Arina h5	varies	uint16	~5 GB each	series demo
haadf/gold_haadf_npy/	NumPy bundle	(4096, 4096)	uint16	~34 MB	workshop / Colab Show2D
haadf/gold_haadf.emd	Velox EMD	(4096, 4096)	uint16	a few MB	full optics carrier

Each _npy* bundle ships a meta.json next to data.npy: shape, dtype, sampling, units, voltage / probe / CL (with provenance flags) when known.

Power-user path (full data, GPU decompression)

Got an NVIDIA GPU and want the full Arina h5 / Velox EMD path? Install quantem.live:

from quantem.live import io
from quantem.widget import Show4DSTEM, Show2D
import torch

folder = io.download("gold_512")
result = io.load(io.discover_masters(folder)[0], det_bin=2)
Show4DSTEM(torch.from_dlpack(result.data))

ds = io.read_image(io.download("gold_haadf"))
Show2D(ds)

Memory (VRAM) for the full h5

det_bin	detector	loaded	peak VRAM	fits 16 GB?
1	192×192	18 GB	~25 GB	no
2	96×96	4.5 GB	~6.9 GB	yes
4	48×48	1.1 GB	~2.2 GB	yes
8	24×24	0.3 GB	~0.5 GB	yes

Licence

CC-BY-4.0. Cite quantem.live / quantem.widget if you use these in a publication.