nayoung10/Packora-data

Packora data

Project links

• —Project page
• —Try Packora (live demo)
• —Paper
• —Code
• —Model checkpoints
• —Hugging Face collection

This repository contains the CSV refcode manifests used to reconstruct the Packora training and benchmark datasets with a licensed Cambridge Structural Database installation. It also provides the aggregate dataset statistics used by the released models. It does not contain CSD structures or prebuilt LMDBs.

csv_manifests/csd.csv is the Packora ablation dataset, previously described as csd_ablations. csv_manifests/csd_clari.csv contains only the published training and validation partitions; benchmark/test inputs are supplied as separate CSV files.

Build datasets

Clone the Packora code and set one common data root containing this repository:

export PACKORA_DATA_ROOT=/path/to/Packora-data

python scripts/convert_csv_to_dataset.py \
  --manifest "$PACKORA_DATA_ROOT/csv_manifests/csd.csv" \
  --output-dir "$PACKORA_DATA_ROOT/csd"

python scripts/convert_csv_to_dataset.py \
  --manifest "$PACKORA_DATA_ROOT/csv_manifests/csd_clari.csv" \
  --output-dir "$PACKORA_DATA_ROOT/csd_clari"

for benchmark in \
  ccdc_teaching csp_blind5 csp_blind6 csp_blind7 flexible rigid; do
  python scripts/convert_csv_to_dataset.py \
    --manifest "$PACKORA_DATA_ROOT/csv_manifests/${benchmark}.csv" \
    --output-dir "$PACKORA_DATA_ROOT/csd_benchmarks"
done

The converter requires compatible CSD/CCDC and RDKit versions. Each split produces <split>.lmdb/, <split>.num_atoms.npy, and <split>.csd_families.npy, plus conversion reports and resumable caches. For benchmarks, the two NPY files are verification and batching caches rather than requirements for ordinary prediction or evaluation: atom counts support filtering or bucketed batching, while families support family-aware sampling.

The released scaler statistics are available at csd/dataset_stats.json and csd_clari/dataset_stats.json. To reproduce or regenerate them after converting the training datasets, run:

python scripts/extract_dataset_stats.py \
  --data_dir "$PACKORA_DATA_ROOT" --dataset_name csd
python scripts/extract_dataset_stats.py \
  --data_dir "$PACKORA_DATA_ROOT" --dataset_name csd_clari

See the Packora project README for the complete expected directory tree, training, prediction, evaluation, and licensing notes. Verify downloaded files with sha256sum -c SHA256SUMS.