neurodeskorg/skills-comm-ground-truth

skills-comm benchmark ground truth

Frozen reference volumes for the automated graders in nipreps/skills-comm. Each grader scores an agent's submission against the reference for its task and returns a verdict plus a 0–100 quality score.

Code lives in the GitHub repository; only the reference volumes are here.

Layout

<category>_gt/<task-id>/<file>.nii.gz

The path matches the ref_dir recorded for each task in benchmark/harness/run_manifest.json, so the harness resolves a reference from the manifest alone.

Fetching

from huggingface_hub import hf_hub_download

p = hf_hub_download(
    repo_id="neurodeskorg/skills-comm-ground-truth",
    filename="structural_gt/structural-brain-extraction/consensus_mask.nii.gz",
    repo_type="dataset",
)

Pin a revision for a reproducible run:

p = hf_hub_download(..., revision="<commit-sha>")

What the references are

Expert manual annotation. clinical-ms-lesion-segmentation (consensus of three independent raters), clinical-stroke-lesion-segmentation (expert tracing shipped with ARC), clinical-wmh-segmentation (WMH Challenge 2017 primary observer mask). Automated lesion tools disagree too much for a tool consensus to be trustworthy, so these grade against the human annotation.

Multi-tool consensus. The brain-extraction tasks, structural-tissue-segmentation-7t and diffusion-brain-mask ship a STAPLE or majority-vote consensus of a curated tool panel, split into core / margin / background zones. consensus_mask.nii.gz is the binary consensus; consensus_zones.nii.gz encodes 0 = background, 1 = margin (unscored), 2 = core (must be kept). structural-brain-extraction-7t-nodura adds dura_envelope.nii.gz, a dura-out pial envelope used by an over-inclusion gate. structural-brain-extraction-stroke also carries the expert lesion_mask.nii.gz used by its lesion-retention gate.

Template prior. structural-mni-registration and functional-bold-to-mni grade against MNI152NLin2009cAsym itself — template_T1w, template_brain_mask and the GM/WM/CSF template_probseg priors — which state where anatomy belongs in that space independently of any registration software. These files are unmodified TemplateFlow volumes.

Sources and terms

Reference volumes are derived from, or shipped with, the datasets below. Terms follow the source; check the source before redistributing or using commercially.

task	source	notes
structural-brain-extraction	OpenNeuro ds002790	derived (tool consensus)
structural-brain-extraction-motion	OpenNeuro ds004173	derived (tool consensus)
structural-brain-extraction-pediatric	OpenNeuro ds000228	derived (tool consensus)
structural-brain-extraction-7t	OpenNeuro ds003642 (CEREBRUM-7T)	derived (tool consensus)
structural-brain-extraction-7t-nodura	OpenNeuro ds003642	derived; envelope from the shipped segmentations
structural-tissue-segmentation-7t	OpenNeuro ds003642	derived (majority vote of shipped segmentations)
structural-brain-extraction-stroke	OpenNeuro ds004884 (ARC)	derived consensus + expert lesion tracing
clinical-stroke-lesion-segmentation	OpenNeuro ds004884 (ARC)	expert manual tracing
diffusion-brain-mask	OpenNeuro ds001226	derived (tool consensus)
structural-mni-registration	TemplateFlow tpl-MNI152NLin2009cAsym	unmodified template files
functional-bold-to-mni	TemplateFlow tpl-MNI152NLin2009cAsym	unmodified template files
clinical-ms-lesion-segmentation	Ljubljana MS database (open_ms_data)	expert consensus mask; source is CC-BY
clinical-wmh-segmentation	WMH Segmentation Challenge 2017 (doi:10.34894/AECRSD)	expert mask; source is CC-BY-NC-4.0, non-commercial

clinical-wmh-segmentation carries the most restrictive terms in the collection: its source is CC-BY-NC-4.0, so that mask must not be used commercially. It was obtained through the open DataverseNL access API, which grants redistribution with attribution under that licence.

TemplateFlow volumes are redistributable under the template's own licence; cite Fonov et al. 2011 and TemplateFlow with any use.

Full per-path terms are in LICENSE.md.

Use in the benchmark

These files are read only by the grading step, never by the agent being evaluated. The benchmark keeps execution and grading separate so the reference is not present where a submission is produced. Anyone using this dataset to evaluate models should preserve that separation, and should be aware that a public reference can enter future training corpora, which would inflate scores without any visible failure.

Citation

Cite the source dataset for the task you use, and the benchmark repository: https://github.com/nipreps/skills-comm.