New TokenPrint method traces language model origins and training data

Researchers have developed a new method called TokenPrint to identify the origin and training data of language models. This technique uses a fingerprint based on the top-k vocabulary projections of late hidden states, compared using Jaccard overlap on decoded token strings. The method demonstrates a "similarity ladder" that correlates with model relatedness, with independently trained models on identical data scoring higher than those with no documented relationship. TokenPrint also functions as a lineage-retrieval tool, accurately identifying base models even when metadata is insufficient, and remains stable under quantization. AI

IMPACT Enables better tracking of model origins and training data, crucial for responsible AI development and governance.

RANK_REASON The cluster describes a new research paper detailing a novel method for language model provenance. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• Jaccard
• language model
• ScienceCast
• TokenPrint

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →