MiniKG

Minimal code for running MiniKG on Family, YAGO3-10, CoDEx-S/M/L, Wikidata5M, and Freebase.

Setup

pip install -r requirements.txt

CUDA and a working C++/CUDA compiler are required.

Data

Family, YAGO3-10, and CoDEx-S/M/L are included as named triples:

data/<dataset>/triples.tsv

Each dataset is one file containing only tab-separated head name, relation name, and tail name triples. Integer IDs are created only in memory.

Download the large datasets from:

• Wikidata5M: https://github.com/THU-KEG/KEPLER#pre-training
• Freebase: https://aws-dglke.readthedocs.io/en/latest/train.html

Prepare them as:

data/wikidata5m/all_id.npy
data/freebase/train.npy

Rows must contain integer IDs in this order:

head relation tail

Run

python run.py --dataset family
python run.py --dataset yago3-10
python run.py --dataset codex-s
python run.py --dataset codex-m
python run.py --dataset codex-l
python run.py --dataset wikidata5m
python run.py --dataset freebase

Training steps are set per dataset in the shared config.json. Results are written to runs/<dataset>/. Large-graph training automatically chooses the shortest frequency-ranked relation prefix covering 70% of Wikidata5M or 80% of Freebase facts. Relation profiling and first-hop decode pruning then use fixed defaults. Small graphs use lightweight compression probes and run the full decode once. Large graphs decode once for profiling and once after focused continuation.

Query

Training also writes a directly queryable representation to runs/<dataset>/query/.

python query.py --dataset family --head 0 --relations 1
python query.py --dataset family --head 0 --relations 1,2

The second command evaluates a two-hop relation path. Answers are returned as entity IDs in JSON.