(ACL 2026 Main) LLMSurgeon recovers the pretraining data mixture of any LLM from only its generated text — no weights, no training data. A calibrated domain classifier plus label-shift correction de-blurs biased predictions. Ships with LLMScan, a benchmark on 8 open-source LLMs.
data-provenance label-shift membership-inference large-language-models llm-analysis pretraining-data data-mixture acl2026
-
Updated
Apr 22, 2026 - Python