PR #1571

open

[Non-Record] Geodesic Topological Tokenizer + BigramHash + BRN

val_bpb
1.5406
Architecture
Transformer
Optimizer
Artifact Size

Training Techniques

Architecture
BigramHash
Hashes consecutive token pairs into a learned embedding table and gates them into the input representation to inject bigram statistics.
parameters: null
depth recurrence
Uses a Belief Recurrent Network (BRN) that maintains a persistent belief state across the sequence and gates representation updates.
parameters: null
Other
other
Geodesic Topological Tokenizer built from persistent homology on n-gram co-occurrence manifolds, using Isomap projection and topologically stable clustering.
parameters: null

Novel Contributions

  • Geodesic Topological Tokenizer based on persistent homology and Isomap
  • BigramHash for injecting learned bigram priors
  • Belief Recurrent Network (BRN) with persistent belief state
  • Reported tokenizer compression advantage of 3.236 chars/token