MODAR: optional-modality autoregressive crystal generation

Prepared on 17 October 2026. This repository contains six trained checkpoints, model configurations and tokenizers. Implementation code is not included in this release. These are custom PyTorch architectures, not Transformers AutoModel checkpoints; the weights alone do not provide a standalone inference program. The authors have not yet selected a license for the model weights. No license from the separate implementation is asserted for these artifacts.

Checkpoints and applications

Each variant uses the optional-modality crystal-language design. The following applications refer to distinct trained checkpoints.

Directory under models/	Parameters	Spectral encoder	Documented application
modar-cnn	309,495,936	Raster CNN	Full simulated modality ladder, paired neutron subset, base size comparisons and clean-input controls
modar-di	385,764,352	Three-layer peak-set transformer	Public experimental PXRD, CNRS/RRUFF and the author-confirmed LiGaCl4 comparison
modar-cnn-noise-adapted	309,495,936	Raster CNN	Noisy-input evaluation and laboratory powder series, including LiCoO2
modar-cnn-large-cell	309,495,936	Raster CNN	Large-cell specialist
modar-cnn-large-cell-noise-adapted	309,495,936	Raster CNN	Noise-adapted large-cell specialist comparison
modar-mp20	131,672,960	Raster CNN	Separate MP-20 evaluation after RealPXRD pretraining and MP-20 fine-tuning

Saved evaluation metadata identifies the raster base model for the full simulated ladder and the paired neutron subset (816 and 280 evaluation shards, respectively). The LiGaCl4 checkpoint identity was confirmed by the authors. Historical evaluation summaries did not contain checkpoint-content hashes; MANIFEST.json identifies the exact files distributed in this repository.

Files and input convention

Every model directory contains:

• checkpoint_last.pt: tensor state dictionary, configuration metadata and step.
• config.json: the same configuration metadata in readable JSON.
• tokenizer.json: the crystal-language token vocabulary.

The checkpoint dictionaries contain model, config and step; no optimizer state or serialized model class is included. Machine-specific paths in the configuration were replaced by portable paths. The numerical tensors were not changed during packaging. The rep and arch configuration identifiers are retained for compatibility with the implementation.

The paired dataset is available at Coooolin/MODAR-data. Its model-facing records contain cond (tokenized condition blocks) and spectra (peak positions and intensities). Reference target strings are evaluation/training labels, not inference inputs. Spectrum positions are 2-theta in degrees, with a standard simulated wavelength of 1.54184 angstrom. The peak-list model converts angles to d-spacings; the raster model uses an angular representation. Formula conditions supply exact stoichiometry, and a supplied lattice is a separate optional condition.

Validation and scope

VALIDATION.json records CPU loading and finite forward checks using the packaged peak-list, base raster and MP-20 checkpoints and compressed dataset inputs. Separate two-token sampling checks also exercised an absent-neutron input. These checks used the implementation retained by the authors; that code is not part of this repository. They verify artifact loading and forward execution, not a new measurement of benchmark accuracy.

The published protocols additionally specify candidate budgets, physical ranking and, where applicable, potential-based refinement. The independent MP-20 checkpoint has a separate pretraining-overlap audit in the dataset release.

Associated work

An optional-modality autoregressive language for crystal-structure solution from complementary X-ray and neutron probes. Manuscript; a final DOI or preprint identifier can be added when available. No publication or acceptance status is implied by these artifacts.