InformedXRD Case-Study Checkpoints
This Hugging Face model repository hosts the NeurIPS artifact release checkpoints associated with InformedXRD.
Contents
This model repo contains only the checkpoints needed to reproduce the paper's evaluation case studies:
- four ViT case-study checkpoints total, including the final mixed champion, Stage-2c and PO comparison checkpoints, and the legacy replay-era baseline
- one 1D residual CNN control checkpoint (submitted epoch-6), plus the validation-selected epoch-46 checkpoint used for the paper's convergence test
- the six XRDBench architecture checkpoints (CNN10, CNN11, PatchTST, BiLSTM,
BiGRU, Transformer) trained under the unchanged public SimXRD
train_1contract and rescored in extinction-group space in the paper; sha256 digests are recorded inmetadata/checkpoint_sha256_manifest.csvand match the digests embedded in the released evaluation artifacts - the matched SG-vs-EG target-comparison pair: the 230-way space-group arm and the 99-way extinction-group arm used for the target-choice study
It does not contain:
- the full companion-paper checkpoint archive
- training code snapshots
- benchmark HDF5 payloads
Recommended filenames
The staged CNN control checkpoint has been renamed to a neutral, paper-facing artifact name:
pxrd_cnn_control_checkpoint.pt
The remaining ViT filenames are kept as-is for now because they align with the existing public checkpoint manifest and reviewer notebooks.
Intended use
These checkpoints are intended for:
- re-running the NeurIPS evaluation protocol on
RRUFF-473andRRUFF-325 - validating the hierarchy-aware error metric
- validating the prior-baseline and stratified-analysis claims
They are not intended as generic pretrained models for production PXRD workflows.
Provenance
The CNN control belongs to the validated in-house Vista lineage documented in:
docs/review_notes/cnn_rruff325_identification_2026-04-29.mdin the private repo
The ViT checkpoints correspond to the released case-study models enumerated in the public checkpoint manifest.
The associated code, manifests, and reviewer notebooks live in the separate anonymous GitHub reproducibility repo:
The associated benchmark payloads live in the separate Hugging Face dataset repo:
Release Status
This repository is intended to be a frozen reviewer-facing artifact. Future corrections, if any, should be versioned explicitly rather than silently replacing the submitted checkpoints.