LLMSurgeon: Diagnosing Data Mixture of Large Language Models

2026-05-28

AI Summary

This work addresses the opacity of pretraining data mixture ratios in large language models, which hinders post-hoc auditing of their corpus composition. It formalizes data mixture diagnosis as an inverse problem under the label shift assumption and introduces the first auditable framework that operates without access to the original training data. By calibrating a soft confusion matrix and solving a constrained inverse problem, the method faithfully reconstructs the domain distribution of the pretraining corpus from model-generated text. Experiments on LLMScan—a newly curated, verifiable evaluation benchmark—demonstrate that the approach accurately recovers the pretraining domain mixture proportions of multiple open-source large language models, offering an effective means to audit their “digital DNA.”

Abstract

The pretraining data mixture of Large Language Models (LLMs) constitutes their "digital DNA", shaping model behaviors, capabilities, and failure modes. Yet this composition is rarely disclosed, making post-hoc auditing of data combination or provenance difficult. In this work, we formalize $\textbf{Data Mixture Surgery (DMS)}$: given only generated text from a target LLM, estimate the domain-level distribution of its pretraining corpus under a predefined taxonomy. We propose $\textbf{LLMSurgeon}$, a strong framework that casts DMS as an inverse problem under the label-shift assumption. Rather than directly aggregating classifier outputs, LLMSurgeon estimates a calibrated $\textit{soft}$ confusion matrix and solves a constrained inverse problem to correct systematic domain confusion and recover the latent mixture prior. To evaluate, we introduce $\textbf{LLMScan}$, a recipe-verifiable evaluation suite built from open-source LLMs with transparent pretraining mixtures. Across LLMScan, LLMSurgeon recovers domain mixtures with high fidelity under fixed protocols. Our work presents a practical, post-hoc approach for auditing the digital DNA of foundation models without access to their training data.

Problem

• Data Mixture Surgery
• pretraining data mixture
• model auditing
• digital DNA
• domain distribution

Innovation

• Data Mixture Surgery
• LLMSurgeon
• inverse problem
• label-shift assumption
• soft confusion matrix