WARP: Weight-Space Analysis for Recovering Training Data Portfolios

2026-07-02

AI Summary

This work addresses the lack of transparency in the domain mixture ratios of training data for foundation models, which hinders understanding of their training distributions. The authors propose a novel method that analyzes the trajectory between a fine-tuned model and its base counterpart in weight space, leveraging model merging to generate pseudo-checkpoints from which geometric features are extracted to infer the global composition of training data. This approach is the first to recover overall domain proportions directly from the geometric structure of weight space, overcoming the limitation of traditional membership inference attacks that can only determine the inclusion of individual samples. Experimental results demonstrate strong performance, achieving mean absolute errors (MAE) of 0.046 and 0.104 on BERT and GPT-2, respectively, outperforming existing baseline methods.

Abstract

Foundation models are routinely released to the public, yet the data recipes used to train them -- such as domain mixture weights that determine how different sources are sampled -- are rarely disclosed. This creates an access asymmetry: researchers study the resulting models but lack visibility into the training distribution that produces them. Prior works for inferring training data, such as membership inference, detect at the level of individual samples and thus cannot characterize the global composition of the training corpus. We introduce WARP, a framework that recovers a fine-tuned model's training mixtures directly from its released weights. WARP interpolates between the base and fine-tuned models using model merging, generating pseudo-checkpoints that approximate the missing training trajectory and expose a geometric footprint of the training data in the weight space. From these simulated footprints, WARP extracts geometric features and maps them to domain proportions using either a parameter-free softmax readout or an MLP projector trained on synthetic mixtures. In controlled experiments with BERT and GPT-2, WARP recovers domain mixtures with an average MAE as low as 0.046 and 0.104 respectively, outperforming membership inference and a variant with access to the true training trajectory.

Problem

• training data recovery
• domain mixture weights
• foundation models
• weight-space analysis
• training distribution

Innovation

• weight-space analysis
• training data recovery
• model merging
• domain mixture inference
• geometric footprint