Skip to content

VASP results to CatBench

Turn your own VASP calculations into a CatBench dataset. The conversion runs on a copy: your original calculations are hashed before and checked after, and never modified.

Ask your LLM

Turn my VASP calculations in ./my_dataset into a CatBench dataset called my_dataset.

What you prepare

Finished calculations with a CONTCAR and an OSZICAR in every directory, laid out like this:

my_dataset/
├── Pt111/                 # one directory per surface
│   ├── slab/              # the clean slab
│   ├── H/                 # one directory per adsorbate
│   └── OH/
└── gas/
    ├── H2gas/             # one directory per gas reference
    └── H2Ogas/

What it asks you

  • the reaction coefficients for each adsorbate — the adsorption energy is adslab − slab − Σ gas terms, e.g. H = adslab − slab − ½ H2. The agent proposes candidates from your gas references; you confirm them;
  • the dataset name.

What you get

  • stage/raw_data/<name>_adsorption.json — ready for Benchmark with CatBench;
  • a check that every original file is unchanged;
  • stage/run_stage.sh — reruns the conversion.

Symlinks, mounted trees and existing files the tool did not create are refused rather than guessed.

Run it yourself 
# 1. scan: complete and incomplete pairs, proposed coefficients
python3 scripts/catbench_vasp_stage.py scan --source my_dataset --out stage

# 2. confirm coefficients in a file outside stage/, e.g. coeff_setting.json:
#    {"H":  {"slab": -1, "adslab": 1, "H2gas": -0.5},
#     "OH": {"slab": -1, "adslab": 1, "H2Ogas": -1, "H2gas": 0.5}}

# 3. stage: copy what catbench reads and write the runner
python3 scripts/catbench_vasp_stage.py stage --source my_dataset --dest stage \
    --dataset-name my_dataset --coeff coeff_setting.json --python <env>/bin/python

# 4. convert on the copy
sh stage/run_stage.sh

# 5. verify the originals and the JSON
python3 scripts/catbench_vasp_stage.py verify --source my_dataset --dest stage
<env>/bin/python scripts/catbench_datasets.py --check stage/raw_data/my_dataset_adsorption.json

<env>/bin/python is any installed model env; every env has catbench.