---
language: en
license: cc-by-4.0
task_categories:
  - other
tags:
  - materials-science
  - x-ray-diffraction
  - crystallography
  - benchmark
  - neurips
---

# InformedXRD Benchmarks

This Hugging Face dataset repository hosts the NeurIPS Evaluations & Datasets artifact release associated with *InformedXRD*.

## Contents

This dataset repo contains:

- the frozen `RRUFF-473` benchmark payload
- the frozen `RRUFF-325` benchmark payload
- a small reviewer subset with compact precomputed inference outputs
- a compact aggregate extinction-group prior for calibration
- manifests and metadata needed to reproduce the benchmark definition

It does **not** contain:

- training code
- the broader PRX archival release
- model checkpoints
- raw ICSD structures, raw diffraction records, or large synthetic training corpora

## Why these benchmarks exist

The paper's main benchmark contribution is not raw scale. It is a reproducible, auditable evaluation surface for real PXRD symmetry classification:

- `RRUFF-473` retains hard real-data cases and supports broad robustness analysis
- `RRUFF-325` is the cleaner downstream subset used for controlled calibration and error-structure analysis

Both benchmarks are tied to explicit curation logic and frozen artifact manifests.

## Intended use

These datasets are intended for:

- extinction-group evaluation on real powder diffraction patterns
- hierarchy-aware rescoring of model predictions
- benchmarking Top-k, prior-baseline, and stratified evaluation claims

The reviewer prior files contain only rounded aggregate extinction-group
probabilities/log-priors and seen-class flags. They do not contain raw ICSD
structures, formulas, atomic coordinates, raw diffraction records, or
per-structure records. The larger trainready corpus used to derive that
aggregate prior is not redistributed.

They are not intended as general-purpose training corpora.

## Croissant metadata

Hugging Face auto-generates core Croissant metadata for supported dataset repos. Because this artifact is distributed primarily as HDF5 plus CSV/JSON sidecars, the release also includes a self-contained Croissant JSON-LD file with core file metadata and Responsible AI fields at `metadata/croissant_rai.jsonld`.

## Companion assets

The associated code, manifests, and reviewer notebooks live in the separate anonymous GitHub reproducibility repo:

- [`neurips-ai-diffraction`](https://anonymous.4open.science/r/neurips-ai-diffraction-4A5D/README.md)

The associated model checkpoints live in the separate Hugging Face model repo:

- [`informedxrd/informedxrd-checkpoints`](https://huggingface.co/informedxrd/informedxrd-checkpoints)

## 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 benchmark payloads.