MathaMAGICal: Validating Trajectory Data Attribution During Capability Acquisition
Jason Lee, Dikshanta Pokhrel, Phonn Nay Khant, Victor Lee, Moritz Weckbecker, Sebastian Lapuschkin, Wojciech Samek, Mark Riedl 等 9 位作者
原文摘要
Influence-function-based methods like MAGIC provide exact first-order scores for trajectory data attribution, yet only retraining indicates whether those scores accurately predict data removal. We evaluate MAGIC on $86$M- and $152$M-parameter GPT-2-style models learning multistep arithmetic from scratch on filtered web text and four synthetic math families. Full-run attribution at $86$M suffers from numerical overflow, and even with metagradient clipping, scores correlate near chance with retraining (linear datamodeling score, LDS, $0.17$--$0.22$). In contrast, short checkpoint windows yield stable, accurate predictions at both scales (LDS $0.86$--$0.99$ across four $100$-step windows at $152$M), whereas a single-checkpoint gradient-dot baseline shows near-zero correlation. Rankings of individual training blocks predict retraining only within that single run and do not transfer across seeds or hyperparameters ($ρ\le 0.09$), whereas format-level patterns replicate across runs. Two data families, worked examples and primitive facts, occupy $2.1$--$2.6\times$ their corpus share in both the helpful and harmful $5\%$ score tails, reflecting strong bidirectional local sensitivity distinct from net benefit under multi-epoch family deletion. These results show that trajectory attribution is reliable within short, validated windows and yields family-level insights that generalize across runs.
cs.LG · cs.AI · cs.CL