Reproduce “STRIDE: Training Data Attribution via Sparse Recovery from Subset Perturbations” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Rishit Dagli, Abir Harrasse, Luke Zhang, Florent Draye, Amirali Abdullah, Bernhard Schölkopf, Zhijing Jin

STRIDE estimates how training examples influence language model predictions by learning activation-space interventions for overlapping training subsets and decoding their effects with sparse regression. On 1.38B-parameter Nanochat pre-trained on Nemotron-ClimbMix, it achieves a Linear Datamodeling Score of 0.1671, versus 0.1284 for GTE. This measures rank agreement with subset-retraining responses; recovery relies on sparsity and approximate additivity.

Key insights

  • The causal impact of training data on LLM behavior can be robustly and efficiently modeled as functional shifts in the activation space, bypassing parameter-space complexities.
  • Training data influence on a specific model prediction is sparse, meaning only a small fraction of the training set significantly contributes to an individual output.
  • By learning zero-shot counterfactual simulators (steering operators) and combining them with a sparse recovery algorithm, it's possible to scale causally-grounded TDA to LLMs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.