dttutty/frost_dataset

FROST DGB Temporal Graph Datasets

This repository packages the curated temporal graph datasets used by FROST as a flat artifact bundle and is published on Hugging Face as dttutty/frost_dataset. Each top-level dataset directory stores the canonical preprocessed edge list plus any optional sidecar arrays needed for downstream experiments. The layout is intended for exact file download with hf download, snapshot_download(), or hf_hub_download(), not for automatic datasets.load_dataset() ingestion or a Viewer-first tabular experience.

The 13 directories listed below are derived from the DGB Zenodo snapshot. Any TGL/TGB artifacts stored in this repository have separate upstream provenance and are not covered by the DGB manifests or protocol claims.

The upstream Zenodo record does not currently specify a value in its License field. The manifests record that fact rather than inventing a license; anyone redistributing this bundle should separately verify the terms of the original networks and the derived artifacts.

Download From Hugging Face

Use the dataset repository as a file bundle:

# Full bundle
hf download dttutty/frost_dataset --repo-type dataset --local-dir DATA

# One dataset directory only
hf download dttutty/frost_dataset --repo-type dataset --include "LASTFM/*" --local-dir DATA

MAG ext_full.npz

MAG/ext_full.npz is not hosted in this Hugging Face repository because the approximately 30 GB artifact would make the private repository exceed Hugging Face's storage quota. Download the artifact directly from the official TGL dataset bucket and place it at MAG/ext_full.npz:

mkdir -p MAG
wget -c \
  https://s3.us-west-2.amazonaws.com/dgl-data/dataset/tgl/MAG/ext_full.npz \
  -O MAG/ext_full.npz

The command assumes that it is run from the root of this repository. The upstream URL is the one published by Amazon Science in TGL's down.sh script.

If you do not want to install hf globally, uvx --from huggingface_hub hf ... works the same way. --local-dir DATA mirrors the repository tree into DATA/ and creates DATA/.cache/huggingface/ metadata for incremental refreshes.

Reproducing And Validating Artifacts

This repository contains published data artifacts only. The transformation and validation code is maintained and tested in the dttutty/FROST repository so that the runtime format and its producers have one source of truth.

From a FROST checkout, use its uv environment and frost-data command:

# Rebuild the reverse temporal CSR into a scratch artifact path.
uv run frost-data build-graph DATA/LASTFM/edges.csv \
  --output artifacts/data/LASTFM/full_graph_with_reverse_edges.npz

# Regenerate a strict timestamp split without dividing tied timestamps.
uv run frost-data split DATA/LASTFM/edges.csv \
  --strategy timestamp --output artifacts/data/LASTFM/edges.csv

# Generate fixed evaluation negatives from real destination nodes only.
uv run frost-data negatives DATA/LASTFM/edges.csv \
  --output artifacts/data/LASTFM/neg_pre_sampling_1_result.csv

# Validate schema, split causality, exact CSR contents, sidecars, and negatives.
uv run frost-data validate DATA/LASTFM

frost-data import-dgb reproduces the complete conversion from an extracted Zenodo DGB ml_* directory, including timestamp handling, feature materialization, split assignment, CSR construction, negative generation, and manifest.json.

Training defaults are also owned by FROST's benchmark presets; they are not dataset metadata and are intentionally not duplicated in this bundle.

Repository Layout

• <dataset>/edges.csv is the canonical preprocessed edge table.
• <dataset>/full_graph_with_reverse_edges.npz is the topology cache used directly by FROST runtime loading.
• <dataset>/edge_features.npy is present when the dataset exposes non-zero edge features. Some datasets keep the original float tensor; selected integer-valued datasets use a downcast materialized copy.
• <dataset>/neg_pre_sampling_1_result.csv stores eid,neg_dst fixed evaluation negatives drawn only from observed destination nodes.
• <dataset>/neg_pre_sampling_1_all_nodes_seed0.csv, when present, stores the corresponding fixed evaluation negatives drawn from the complete contiguous node-ID range with random seed 0.
• <dataset>/manifest.json records provenance, split, graph, feature, and negative-sampling policies.

Fixed Evaluation Negative Variants

Both negative files contain one eid,neg_dst row for every row in edges.csv. They corrupt only the destination of the positive edge, exclude that edge's positive destination, and are intended for reproducible evaluation rather than training. They differ in the candidate population used to draw neg_dst:

Artifact	Candidate population	FROST configuration
neg_pre_sampling_1_result.csv	Node IDs that occur in the dst column of edges.csv	neg_sampler=rnd_dst_then_fixed
neg_pre_sampling_1_all_nodes_seed0.csv	Every node ID in the complete contiguous range, including nodes that occur only as sources	neg_sampler=rnd_dst_then_fixed_all

The all_nodes suffix describes the wider candidate population, while seed0 records the random seed used to materialize the fixed samples. Results from the two files use different evaluation protocols and should not be compared as though they were produced from the same negative candidate set.

Use From Python

Use this repository as a file bundle:

from pathlib import Path

from huggingface_hub import snapshot_download

repo_dir = Path(
    snapshot_download(
        repo_id="dttutty/frost_dataset",
        repo_type="dataset",
        local_dir="DATA",
    )
)

edges = repo_dir / "MOOC" / "edges.csv"
graph = repo_dir / "MOOC" / "full_graph_with_reverse_edges.npz"
edge_features = repo_dir / "MOOC" / "edge_features.npy"

If you need a datasets library dataset or a Hub Viewer-backed table, export the per-dataset artifacts into a supported viewer-first layout such as CSV or Parquet with explicit split or config metadata. The current repository mixes CSV, NPY, and NPZ sidecar files and is optimized for artifact download instead of viewer-native browsing.

Format Notes

• Upstream raw DGB networks are originally stored as <dataset>.csv, with one edge per line.
• The raw edge-list schema is source_node,destination_node,timestamp,edge_label,edge_features....
• This bundle keeps the baseline-friendly preprocessed files directly under each dataset directory.
• edges.csv stores the preprocessed event table with columns eid, src, dst, time, and default_split.
• default_split targets 70/15/15 by event count but moves each boundary to a timestamp transition, so one timestamp never spans two splits.
• The original state/classification label column is not retained because FROST's link-prediction runtime does not consume it.
• edge_features.npy stores the dense edge-feature matrix when the dataset contains non-zero edge features.
• Preprocessed .npy files often have one extra leading row for index alignment or padding, so their first dimension is usually edge_count + 1 or node_count + 1.

Normalization Notes

• The published bundle already includes the normalization that FROST expects at runtime.
• CanParl, UNtrade, and UNvote use contiguous 0-based yearly timestamp indices.
• Selected integer-valued edge-feature arrays are materialized with lossless downcasts for storage efficiency.
• Fixed negatives are evaluation-only destination corruptions. Training uses a fresh random destination sample and should not reuse the fixed file every epoch.

Dataset Catalog

Exact edge counts, split sizes, evaluation batch ceilings, graph direction, and artifact shapes are generated from the per-dataset manifests in DATASET_CATALOG.md. Regenerate it from a FROST checkout:

uv run frost-data catalog /path/to/frost_dataset \
  --markdown-output /path/to/frost_dataset/DATASET_CATALOG.md --force

Training defaults remain in FROST's benchmarks/specs/presets.yaml; the generated maximum is only the smaller actual validation/test split.