REIFM checkpoints — Reification as a Transferable Vocabulary

Checkpoints of the paper Reification as a Transferable Vocabulary: Zero-Shot Link Prediction with Vanilla GNNs (Camille Pradel, Matr, 2026; arXiv:2609.11347, https://arxiv.org/abs/2609.11347). Code and evaluation pipeline: https://github.com/getorbital/REIFM.

All models read reified graphs (every fact is a node linked to its subject, object and relation-type node through six meta-relations; relation types are anonymous shared nodes) and are applied zero-shot to unseen graphs: no fine-tuning, no target-side adaptation, no entity or relation embeddings.

Files

directory	model	training data	selection	notes
gat_seed{0,1,2}/	GAT, PyG GATConv (4 heads, edge features), d=64, L=12	FB15k-237-Inductive v1 train (4,245 triples)	validation MRR on the training graph's own validation split (Tier-1)	headline model of the paper (Tables 2–4)
ginemax_seed{0,1,2}/	GINE variant with mean+max aggregation (reifm.models.ReifiedMeanMaxConv), d=64, L=12	same	same
gine_sum_seed{0,1,2}/	GINE, PyG GINEConv, sum aggregation, d=64, L=12	same	same
sage_seed{0,1,2}/	GraphSAGE, PyG SAGEConv (no edge features), d=64, L=12	same	same
rgcn_seed{0,1,2}/	R-GCN, PyG RGCNConv over the 6 meta-relations, d=64, L=12	same	same
corpus_recipe_1kg_seed{0,1,2}/	same architecture and recipe as the generic model, trained on FB15k-237-Inductive v1 alone	FB15k-237-Inductive v1 train	Tier-1 (FB15k-237-Ind v1 validation)	the "ours, 1 KG, corpus recipe" column of Appendix D
generic3kg_seed{A,B,C}/	GINE mean+max (memory-lean implementation ginemaxlm), d=64, L=12, query-conditioned readout	FB15k-237 + WN18RR + CoDEx-Medium (transductive)	mean validation MRR of FB15k-237-Ind v1, WN18RR-Ind v1, NELL-995-Ind v1 (Tier-2, validation graphs distinct from the training graphs, never a test split)	the generic model of the KG→RDB probe (Section 6 / Appendix D); seed A is the original run, B and C re-pretrainings; a fourth draw failed the pretraining sanity gate and is not released as a probe model

Every directory holds best.pt (state dict) and results.json (the full argument block, per-epoch training history, validation-selection trace).

Training recipe (15 KG models)

Dimension 64, 12 layers, dropout 0.2, cosine schedule from 5·10⁻⁴, weight decay 0.01, batch 16 queries, full-graph propagation, plain readout (an MLP over the candidate state), cross-entropy over all entities with known answers masked; 20 epochs capped at 1,800 s on one NVIDIA A100 80 GB (the cap binds for the GAT only: 14 epochs). Seeds {0, 1, 2} fixed a priori; no seed was rejected or replaced. Only backbone and seed differ across the 15 runs (results.json → args).

Expected numbers

Filtered MRR, zero-shot, mean over the 3 seeds (paper Table 2):

model	inductive-(e) (12)	inductive-(e,r) (13)	extended (15)	all 40
GAT	0.5457	0.3300	0.2281	0.3565
GINE (mean+max)	0.5276	0.3148	0.1807	0.3284
GINE (sum)	0.5069	0.2918	0.1987	0.3214
GraphSAGE	0.4676	0.2801	0.1625	0.2923
R-GCN	0.4344	0.2005	0.1055	0.2350
ULTRA-3g (local re-run, reference)	0.5224	0.3446	0.2786	0.3732

Per-seed and per-split values: results/broad_eval_matrix.csv in the code repository. To reproduce one cell:

uv run python scripts/eval_ckpt.py checkpoints/gat_seed0/best.pt \
    --eval WN18RRInductive:v1 --backbone gat --dim 64 --layers 12

Loading is strict (load_state_dict(strict=True)): the architecture flags must match the ones in results.json.

Known limitation

On FBNELL (the one benchmark of the 40 whose inference graph contains isolated entities), all 15 KG models collapse to MRR 0.08–0.11 with Hits@1 = 0: the five isolated entities, unreachable by propagation and identical under the plain readout, preempt ranks 1–5 of every query. The paper analyses this in Section 5.3.

License

MIT. Training data: FB15k-237-Inductive (GraIL), FB15k-237, WN18RR, CoDEx-Medium, downloaded from their original sources through ULTRA's dataset loaders; not redistributed here.