Until now Haishool knew its facts as separate cards. Version 4b links all 3,081 of them into one graph: Turkey reaches Led Zeppelin, a spoon reaches the Beatles, in a few hops. Where people disagree, the link says who holds it. The model writes every path itself, and every step is checked against the real links.
Dense form, as the model learned it: q turkey hop led_zeppelin. a turkey coast_on mediterranean_sea coast_of morocco visited_by robert_plant member_of led_zeppelin.
A model is a mirror of the culture that describes the world. Contested things are not flattened into one "truth": the model learns q jerusalem capital_of. a israel palestine contested. and, for each side, what that side holds.
The links come from four places: 122 written by hand with sources and views; 5,075 where a record names another record (spoon made_of metal); shared values such as europe, 1970s or kitchen; and a few rules (located_in, shares_food) that follow from the others. The model also learns to say no: q turkey borders colombia. a no.
| Version | Trained facts recalled | Unseen facts guessed | Unseen paths: real links | Unseen paths: shortest |
|---|---|---|---|---|
| v3 · separate cards | 100 % | 21.8 % | – | – |
| v4 · 52 linked things | 100 % | 21.3 % | 99 % | 96 % |
| v4b · everything linked | 99.5 % | 35.0 % | 89 % | 66 % |
Same model size (15.8 million parameters), 24,000 training steps on one RTX 4090. "Unseen" means held back from training in a separate run. Linking the facts raised the share of unseen facts the model guesses exactly from 22 % to 35 %.
✓ the link exists · ✗ the model made it up · ◐ held by one side. Questions that are not about links go to the normal Haishool chat.