---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- materials
- xrd
- diffraction
- crystal-structure
- inverse-design
- jarvis
pretty_name: DiffractGPT-DFT (XRD to crystal structure)
---

# DiffractGPT-DFT

Alpaca-format dataset for the **inverse powder-XRD problem**: given a chemical
formula and the main 2-theta peaks of a Cu-K-alpha pattern, generate the crystal
structure. **73,374 train / 8,152 validation** structures from JARVIS-DFT
`dft_3d`. Same text format as [`knc6/diffractgpt_jarvis_dft`](https://huggingface.co/datasets/knc6/diffractgpt_jarvis_dft),
rebuilt on a fixed split used to benchmark multiple base LLMs for XRD-to-structure.

## Record format (4 fields)
| field | content |
|---|---|
| `instruction` | `Below is a description of a material.` |
| `input` | formula + top-20 XRD peaks `<2theta>deg(<rel intensity>)` + a request for the structure |
| `id` | JARVIS identifier |
| `output` | `a b c` (A) / `alpha beta gamma` (deg) / one `Element x y z` line per atom |

Models: https://huggingface.co/knc6/diffractgpt_dft_models ; code:
https://github.com/atomgptlab/diffractgpt ; underlying data: JARVIS-DFT
(https://doi.org/10.1038/s41524-020-00440-1).