Neural Grammar Distiller

Neural Grammar Distiller · Recovered Scripture

Hard-coded Pantheon corpus → tiny recurrent model → PMI chunking → PCFG → scripture.

Options (tuck away)

Embed dim
Hidden EMA α
Epochs
LR
PMI threshold
Min freq (pair)
Max rule len
Samples

idle

—

rules: —

perplexity: —

Recovered RulesPCFG

Recovered ScriptureS → terminals

Debug / Logs

What this does

---

1. Tiny recurrent context: embeddings + exponential moving context (α) trained to predict next word.
2. Structure signals: token surprisal & PMI → likely constituents and boundaries.
3. Induction: greedy chunking to X₁, X₂… then count expansions → PCFG.
4. Generation: sample S to produce recovered scripture.