Skip to content

Project 1: Tokenizer stress-test on medical jargon

Mentors: To be announced

Problem: General-purpose tokenizers weren't built for medical vocabulary, and there's no measurement of how much token overhead drug names, dosages, and ICD codes actually cost compared to plain English.

Context: Built on mtsamples medical transcription reports and CMS ICD-10-CM code descriptions, anchored to ML4LLM Ch.2 · proj1: Three tokenization schemes (helper).

Goals: How many extra tokens does a general-purpose tokenizer burn on drug names, dosages, and ICD codes compared to plain English of the same length?

Deliverables: A notebook that runs the same mtsamples sentences through several tokenization schemes (word-level, BPE, WordPiece), compares token counts on drug-name/dosage spans against matched plain-English spans, and plots token-per-character ratio by category (drug name, dosage, ICD code, plain prose) to show where general-purpose tokenizers fragment medical vocabulary most.

Showcase: TBD

References:

← Back to all Basic Science projects