标签
This paper investigates whether released LLM tokenizer vocabularies can support fine-grained token-level estimation of hidden pretraining corpora, proposing a Quantile-Guided Density Estimation (QGDE) method that achieves low error rates in controlled and realistic settings.