anima-xbind-g1-crack — G1 recombination CRACK (corpus×task-class measure-swap)

303M CLMConvMoE trained on the XBIND synthetic corpus (concept polarity XOR task) — the earned-terminal's identified exit: swap the learning measure (corpus/task class) rather than the architecture. CRACK confirmed: both main seeds held-out D-acc = 1.000, shuffle-control = 0.515 (V-B pass, no leak), Δcontrol = 0.485, C-rate = 1.0. Proves the G1 recombination wall's true cause is the corpus×CE measure — a corpus carrying held-out compositional (XOR) signal makes 303M byte-LM learn held-out recombination perfectly. Card H_9266.

  • xbind_s7.clm / xbind_s4302.clm — main arms (seed 7 / 4302), held-out D-acc 1.000
  • xbind_shuf_s7.clm — shuffle control, held-out D-acc 0.515 (chance)
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support