Expand description
The shared option-scorer: ONE linear projection over anchor-token hidden states, with per-primitive readouts. This is the generalist head (P1.1) — no per-question parameters, no cross-question conditioning: each question is scored exclusively from its own anchor positions, so runtime-declared questions work zero-shot and the 255-option cardinality ceiling costs nothing in parameters.
Structs§
- Question
Score - The scored output for one question: the raw per-anchor logits plus the primitive’s distribution readout, derived on demand.
- Shared
Scorer - One shared linear scorer over anchor-token hidden states: a single
d_model → 1projection applied to every anchor position, of every option, of every question. Trained weights live in a [nn::VarStore] so checkpoints attach to the converted backbone in P1.2 without a format boundary.
Functions§
- soft_
target_ kl - KL divergence
KL(target ‖ p)between a target distribution (the corrected teacher-average histogram in P1.4 distillation) and the scorer’s readout for one question. Distribution targets are the native input of the binned-softmax parameterization (S6b2); the target must be non-negative and sum to 1 over the question’s label arity (2 for noul:[P(false), P(true)]).