--- language: - en tags: - materials-science - crystal-structure - powder-diffraction configs: - config_name: tier-a default: true data_files: - split: train path: data/tier-a/train-*.jsonl.gz - split: validation path: data/tier-a/valid-*.jsonl.gz - split: test path: data/tier-a/test-*.jsonl.gz - config_name: mp20 data_files: - split: train path: data/mp20/train-*.jsonl.gz - split: validation path: data/mp20/valid-*.jsonl.gz - split: test path: data/mp20/test-*.jsonl.gz --- # MODAR paired diffraction and crystal-language data Prepared on 17 October 2026. The authors have not yet selected a release license for their derived data. Original database and collaborator provenance is retained; see `SOURCES.md`. This repository contains data and documentation, without the project implementation. No code license is asserted for these data. ## Contents | Configuration | Train | Validation | Test | Scope | | --- | ---: | ---: | ---: | --- | | `tier-a` | 3,456,202 | 34,039 | 33,944 | Paired PXRD, simulated ND, composition conditions and crystal targets | | `mp20` | 27,136 | 9,047 | 9,046 | Separate official MP-20 split used by the auxiliary model | The frozen source splits are preserved. Each original JSONL is split only at record boundaries, into gzip files containing at most approximately 1 GB of uncompressed data. Concatenating the decompressed shards reproduces the original file byte for byte. `SPLITS.json` records counts and original-stream SHA-256 values; `MANIFEST.json` records each distributed file's size and SHA-256. Additional directories contain: - `benchmarks/`: prepared model inputs for CNRS-126, RRUFF-269, public-337 and the ten laboratory powders. These include reference targets for evaluation, but inference must use only the selected conditions and spectra. - `source_data/`: main-figure numbers, selected supplementary figure data and all 21 numbered supplementary tables (tables S6 and S15 have two parts). - `audits/`: source-deduplication summary and MP-20 pretraining-overlap records. - `model_applications.json`: checkpoint roles, kept separate from the data split. Table S19 corresponds to the accompanying manuscript and distinguishes the checkpoint applications using the saved evaluation records. Experimental recovery values are unchanged. Figure 7 files use the current panel order: public-337 (a), RPS comparison (b), LiCoO2 (c), and LiGaCl4 Top-1 / Top-20 candidate count / Li error (d–f). Published RPS integer counts are approximate conversions from the reported rates; the original percentages are also supplied. The full RealPXRD pretraining corpus and per-query pools of 100 generated candidates are not included in this minimum release. The LiGaCl4 files contain the author-provided repeated-run statistics; the raw POLARIS pattern and generated structures for those runs were not present in the available project assets and are not claimed to be included. ## Record schema | Field | Meaning | | --- | --- | | `id`, `source` | Source database identifier and collection | | `split` | Frozen source split (`train`, `valid` or `test`) | | `n_sites` | Reference primitive-cell site count | | `target`, `sites_target` | Crystal-language target strings; evaluation/training targets, not inference inputs | | `cond` | Token blocks for formula, lattice, elements, elemental mass fractions and auxiliary valence conditions | | `spectra` | `pxrd_x`, `pxrd_y`, `neutron_x`, `neutron_y` peak arrays | | `rep` | Serializer identifier retained for code compatibility | Peak positions are 2-theta in degrees under the stated wavelength convention, not d-spacings. The standard simulated convention is 1.54184 Å over 5–80°. The d–I model converts positions to d-spacings; the raster model uses a 1,501-bin angular representation. Diffraction values in the Tier-A/MP-20 inputs are simulated. Auxiliary ND and composition for the public experimental and laboratory PXRD collections are derived from paired reference structures. LiGaCl4 uses simulated PXRD and separately absent/measured/simulated ND in its repeated comparison. ## Read the data The records are gzip-compressed JSONL. With the optional `datasets` library: ```python from datasets import load_dataset data = load_dataset("Coooolin/MODAR-data", "tier-a", split="test") ``` Alternatively, use standard gzip and JSON readers. To reconstruct an original split, decompress its shards in the order listed in `SPLITS.json` and concatenate their bytes. The resulting SHA-256 must equal the recorded `source_sha256` for that split. Project inference and training code is not included in the model or dataset artifacts. ## Evaluation scope The Tier-A full-test rates refer to 33,944 structures; paired neutron intervals use a 10,000-structure subset. In the auxiliary MP-20 protocol, 6,399 of 9,046 test structures have a structural near-duplicate in the RealPXRD pretraining library under the documented audit. Its recovery rates and candidate budgets are reported separately. Source and input-condition distinctions are essential when reusing the benchmarks.