Sign InOpen Brain
arXivPaperNeeds Review

Measuring Task-Agnostic Training Data Influence Across Language Model Pretraining

A task-agnostic influence measure tracks which examples steer pretraining toward final parameters without choosing a downstream eval. It reveals a literature-to-STEM shift across training.

arXiv · Aug 13, 2026
Open Source Open MarkdownOpen JSON
Source Summary

The method defines an example’s influence by how much its gradient update reduces squared distance to the run’s final parameters. It estimates that value from intermediate checkpoints **without retraining** or selecting a downstream task.

Practical Implication

Across **18 Pythia and PolyPythia configurations**, influential data changed over time: literature-related examples aligned more strongly early, while STEM data aligned more strongly later. Training-pipeline builders can use this trajectory view alongside task-specific attribution when inspecting data mixtures.

Agent-Ready Context
The method defines an example’s influence by how much its gradient update reduces squared distance to the run’s final parameters. It estimates that value from intermediate checkpoints **without retraining** or selecting a downstream task.

Across **18 Pythia and PolyPythia configurations**, influential data changed over time: literature-related examples aligned more strongly early, while STEM data aligned more strongly later. Training-pipeline builders can use this trajectory view alongside task-specific attribution when inspecting data mixtures.

The crossover was qualitative and broadly consistent, but the measure is tied to the final parameters of a particular run. It does not show which examples improve a coding agent, nor replace downstream capability and reliability evaluations.
Connected Context · Feed7 Judgment

This adds a task-agnostic, checkpoint-based view of which pretraining examples most shaped a particular run, complementing candidates focused on downstream behavior and benchmark design. The changing literature-to-STEM influence pattern argues against treating data contribution as static, but the run-relative measure does not establish agent capability or reliability and therefore remains a diagnostic alongside, not a substitute for, task and workflow evaluations.

LittleLearner: Language Models Under Pedagogically Controlled Knowledge ExposureLittleLearner controls prior exposure to study knowledge acquisition, while this method retrospectively estimates how individual examples influenced final parameters; together they offer complementary experimental and diagnostic views of learning from data.State of Data — Sean Cai, Independent / State of DataThe workflow-trace candidate emphasizes scaffold-dependent downstream behavior, whereas this Signal measures pretraining influence without choosing a downstream task; the two operate at different layers and neither substitutes for the other.LACUNA: A Testbed for Evaluating Localization Precision for LLM UnlearningLACUNA supplies ground-truth weight locations for injected information to test unlearning, while this Signal estimates example influence relative to final parameters without such ground truth; LACUNA therefore highlights a validation advantage absent from this broader attribution method.
Context Map
benchmarkresearchdata#benchmark-integrity
Uncertainty
The crossover was qualitative and broadly consistent, but the measure is tied to the final parameters of a particular run. It does not show which examples improve a coding agent, nor replace downstream capability and reliability evaluations.