N00beroUno/custodial-weights-kinship

Custodial Weights — Kinship Dataset

A converted and expanded version of the public kdkyum kinship graph (1000 synthetic families) used to train the Custodial Weights mechanism.

What it is

The base graph is public-domain synthetic kinship data: surnames and given names, with father/mother/son/daughter/brother/sister/husband/wife relations across 1000 families. We convert it to this repo's family-triple format and add three synthetic layers on top:

1. Walk expansion (~4×) — triples stitched into left-to-right narrative walk sequences that share people across relations, so a model learns to compose relations the way a narrative carries them. Deterministic per seed.
2. Contrast negatives — for each probe fact S p O, same-family candidates confirmed by the Gate to not be the true object. These gate-rejected lies are what contrast training drills against.
3. Refresh families — a held-out family set to test adding facts after training without forgetting old ones.

Provenance

• Base data: kdkyum kinship graph (public domain). We are not the authors of the base family names or triples; any use must carry the source dataset's attribution in addition to this card.
• The walks and contrast negatives are generated (not scraped), deterministic per seed.

Files

file	contents
data/family_graph_1000.json	converted family triples, {family_id: {people, triples, direction}} (~4 MB)
data/walk_1000.json	~1.47 GB of walk-expanded training sentences
data/*_refresh*.json	held-out refresh families

The walk file is large; download only what you need. The conversion code is src/scale_data.py in the parent Custodial Weights repo.

Why it's here

It is gitignored from the code repo because it is derivable, but derived datasets are exactly what someone else may want to reuse. This is a genuine contribution — if you use it, clone and build on it.

License

MIT, with the caveat that you remain responsible for carrying the base kinship dataset's attribution and terms in any downstream use.