LEDGER — RelBench checkpoints
Learned Event Distributions for Generic Entity Readout
Checkpoints behind the LEDGER entry on the
RelBench leaderboard.
A relational database is read as a ledger of timestamped events, and one
self-supervised objective — the sufficient statistics of the window
(t_q, t_q + D] — covers every entity task. The task enters only at readout.
No task labels are used in training.
Code: https://github.com/ShantanuAnant/ledger
| Board | Metric | LEDGER | Position |
|---|---|---|---|
| Recommendation | MAP | 8.75 | 3 of 7 |
| Classification | AUROC | 72.89 | 7 of 13 |
| Regression | NMAE | 0.3391 | 8 of 14 |
All 31 tasks scored by python -m relbench.submit (relbench 3.0.1).
Layout
recommendation/ WHO head → link-prediction tasks
window/ WINDOW head → classification + regression tasks
MANIFEST.md task → checkpoint, all 31 tasks
Files are named {dataset}__{entity}.pt. One checkpoint often serves several
tasks; MANIFEST.md is the authoritative mapping.
Usage
import torch
from ledger.model.load import load_checkpoint # from the GitHub repo
ckpt = torch.load("recommendation/rel-amazon__customer.pt",
map_location="cpu", weights_only=False)
print(ckpt["args"]) # the exact training configuration
Evaluate with scripts/eval_rec.py (recommendation) or scripts/eval_entity.py
/ scripts/icl_readout.py (classification, regression). Both need a corpus
built at the matching cutoff — see the repository README.
Note
These are slimmed for inference: model and args only, with optimizer
state removed. They score identically to the originals and can be fine-tuned
from, but training cannot be resumed from them.
Trained on a single NVIDIA B200 (183 GB), CUDA 12.8, PyTorch 2.11.