AI
OriginBlame: AIトレーニングデータセットのレコードレベルおよびトークンレベルのデータ来歴
OriginBlame: Record- and Token-Level Data Provenance for AI Training Datasets
arxiv2026年7月16日
日本語要約
本論文は、AIトレーニングデータセット内のレコードレベルおよびトークンレベルでデータ来歴を追跡するシステムOriginBlameを紹介する。これはAIモデルの完全性、監査可能性、再現性を確保するために不可欠である。信頼性が高く、制御可能なAI開発の基盤を築く。
English Summary
This paper introduces OriginBlame, a system for tracking data provenance at both record and token levels within AI training datasets. This is vital for ensuring the integrity, auditability, and reproducibility of AI models. It lays the groundwork for more trustworthy and controllable AI development.