lexuanbach/score-diffusion-predictions

Score diffusion: cached classifier predictions

Public prediction cache for When Does Score Diffusion Help Graph Conformal Prediction? by Xuan-Bach Le.

These files contain logits, split indices and training metadata. They are not model weights or loadable trained-model checkpoints. The original training script did not save a state_dict. They support recalibration and evaluation on the stored graphs, but cannot run inference on new graphs.

Code, data, supplementary PDF and reproduction instructions: https://github.com/lexuanbach/score-diffusion-artifact

Contents

predictions.tar.gz contains 330 .pt files under results/preds/: 13 benchmark graphs × 5 classifiers (GCN, GAT, GraphSAGE, GCNII, MLP) × 5 seeds, plus 5 GCN temporal-split caches for ogbn-arxiv. Each dictionary has logits, train_idx, val_idx, and info. info identifies the dataset, model, seed, split and training summary.

Archive SHA-256: 5f7d44436735100afe708ecc31efdaca7372d736fe3ffed4a08814c2ee8777b5. predictions-files.json records the size and SHA-256 of every extracted file. This is the same archive published in GitHub release v1.0.0.

Download into the artifact

From the cloned artifact directory:

curl -fL https://huggingface.co/datasets/lexuanbach/score-diffusion-predictions/resolve/main/predictions.tar.gz -o predictions.tar.gz
python - <<'PYCODE'
import hashlib, tarfile
from pathlib import Path
p = Path('predictions.tar.gz')
assert hashlib.sha256(p.read_bytes()).hexdigest() == '5f7d44436735100afe708ecc31efdaca7372d736fe3ffed4a08814c2ee8777b5'
with tarfile.open(p) as archive:
    archive.extractall('.', filter='data')
PYCODE
python scripts/verify.py --bundles

The final check also requires the benchmark data bundle described in the artifact README. Refer to that README for pinned dependencies and commands to rerun conformal evaluation. Full retraining is a separate reproduction level and can differ across platforms.

Scope and attribution

This repository contains research benchmark model outputs only. It contains no manuscript, LaTeX source, review reports or audit reports. Dataset sources and their terms are documented in the artifact's DATA.md and THIRD_PARTY.md. No new license over upstream datasets is implied.