Data access

Dataset selection

Select by domain, geometry and numerical method

Client setup

Install the Hugging Face command-line client

The hf command is provided by huggingface_hub. The optional hf_xet package supports repositories stored with Xet.

python -m pip install -U huggingface_hub hf_xet
hf auth login

The datasets are public. Authentication is still useful for reliable Hub access and higher rate limits; keep tokens out of scripts and repositories.

Repository inspection

Estimate a transfer with --dry-run

A dry run lists matching files and reports the estimated transfer without downloading them.

hf download neashton/ahmedml \
  --type dataset \
  --local-dir ./ahmedml_data \
  --dry-run

Replace the repository ID with neashton/windsorml, neashton/drivaerml or nvidia/HiLiftAeroML. Each dataset page provides a builder for its actual file patterns.

File selection

Select data products explicitly

hf download neashton/ahmedml \
  --type dataset \
  --local-dir ./ahmedml_data \
  --include "force_mom_all.csv" \
  --include "geo_parameters_all.csv" \
  --include "run_*/ahmed_*.stl" \
  --dry-run

1. Tabular dataConsolidated CSV files contain geometry parameters and integrated coefficients.
2. GeometrySTL or STEP files provide geometric input for shape-based models.
3. Surface fieldsVTP or VTU boundary files contain spatially resolved surface quantities.
4. Volume fieldsThree-dimensional field files have the largest storage and processing requirements.

Transfer

Review the estimate before removing --dry-run

• Confirm the reported byte count and number of files.
• Allow additional working space for compressed archives and derived training data.
• Use a persistent filesystem; large transfers may need to resume.
• Record the repository revision or commit hash used by an experiment.
• Review dataset-specific known constraints before creating splits.