Tutorial · 2026-07-01
DNA tokenization and safety classes
A deeper look at A/C/G/T strings, k-mers, BPE-style genome tokenization, and computational-only safety boundaries.
Students and researchers learning how Hermon DNA differs from wet-lab protocol generation.
01
From bases to tokens
DNA sequences can be represented as bases, overlapping k-mers, non-overlapping segments, or learned tokens. The representation changes which downstream tasks are efficient.
02
Encoder tasks
DNABERT-style encoders fit classification, embeddings, search, and sequence linting. They are not a reason to let a public model produce experimental protocols.
03
Safety classes
Hermon DNA should classify requests as computational explanation, symbolic simulation, needs expert review, denied wet-lab protocol, or denied harmful biological design.