---
license: mit
library_name: pytorch
tags:
  - crystallography
  - x-ray-diffraction
  - point-cloud
  - single-crystal
  - sidit
---

# SiDIT base

This repository contains the tensor-only parameters of the historical SiDIT model for joint reflection-family and crystal-property prediction from single-crystal diffraction point clouds.

The inference implementation and numerical preprocessing are available at [max6616/SiDIT on GitHub](https://github.com/max6616/SiDIT). Recovered test inputs are distributed in [max6616/SiD3M-test](https://huggingface.co/datasets/max6616/SiD3M-test); consult the dataset card for their exact scope.

## Parameters and configuration

`sidit-base-state-dict.pt` contains 523 state tensors, 194,915,607 bytes in total. It preserves the model tensors associated with the retained cached benchmark and excludes optimizer state and Python argument objects. Loading uses `torch.load(path, map_location='cpu', weights_only=True)`; construct the model with the accompanying `config.json` and the GitHub implementation.

- Backbone: Point Transformer V3 encoder–decoder, base configuration.
- Per-point classes: 21 classes for each of h, k, and l, corresponding to `[-10, 10]`.
- Global outputs: crystal system, Laue class, point group, space group, and the nine-component cell head.
- Cell head: normalized lengths plus cosine/sine angles; decode in float32 as documented in the implementation.
- The retained checkpoint was saved at global step 80,000. Its tensors are unchanged by this release.

Model-file SHA-256: `22fceed0e9e303e97ce668cf71b092f39ab9446fe38ec350a4189bc97e389ad7`.

## Scope

The parameters support inference on the supplied simulated point-cloud format. Predicted HKL values are canonical reflection-family labels in that format. The model does not supply a refined orientation matrix or establish an end-to-end experimental indexing result. Public test recovery and cached statistical reconstruction are identified separately in the dataset card.

Authors: Zhao Zhang, Zheng Dong, Zhi Geng, Xinlong Dong, Yi Zhang, and Gaoqi He. Original implementation copyright 2025 Zhang Zhao; MIT license. Pointcept-derived code retains its upstream MIT notice in the software repository.