Data Mixture Inference: What do BPE Tokenizers Reveal about their Training Data?

Name: Data Mixture Inference: What do BPE Tokenizers Reveal about their Training Data?
Uploaded: 2026-01-27T00:00:00.000Z
Duration: 44 min 29 s
Description: This talk presents the paper *Data Mixture Inference: What do BPE tokenizers reveal about their training data?* The core result is that a BPE tokenizer's ordered merge list leaks the proportions of languages, code and...

Google TechTalks · 44:29

This talk presents the paper *Data Mixture Inference: What do BPE tokenizers reveal about their training data?* The core result is that a BPE tokenizer's ordered merge list leaks the proportions of languages, code and...

Read the full summary on tuber