Research series

Training Data Attribution for NLP and LLM Research

A thesis- and interview-oriented path through attribution units, utilities, Shapley values, ablation, causal evidence, instance-level methods, scaling, uncertainty, validation, and research extensions.

Start the series
中文
Attribution category

Part I

Define the attribution problem

  1. 1. What Is Training Data Attribution?
  2. 2. Attribution Units: From Data Sources to Individual Examples
  3. 3. Utility Functions: Which Model Behaviours Are We Explaining?

Part II

Contribution under coalition context

  1. 1. Coalitions, Marginal Contributions, and Shapley Values
  2. 2. Ablation vs Shapley: Why Coalition Context Matters
  3. 3. From Association to Causal Evidence: What Can Attribution Claim?

Part III

Methods, scaling, and uncertainty

  1. 1. Instance-Level Attribution: Gradient Similarity, Influence Functions, and TracIn
  2. 2. Scaling Data Attribution: Monte Carlo Shapley, Surrogate Models, and Hierarchical Attribution
  3. 3. Statistical Reliability: Bias, Variance, Bootstrap, and Attribution Uncertainty

Part IV

Validation and interview narrative

  1. 1. Data Intervention and Attribution Validation
  2. 2. Interview Notes: Explaining My Training-Data Attribution Thesis

How to use this series

The goal is not to sound like a generic tutorial. The goal is to make my thesis explainable, defensible, and expandable.

Interview readiness

Each note ends with a spoken-style answer that can be adapted for project interviews and PhD discussions.

Research precision

The series separates attribution units, utilities, estimators, interventions, and uncertainty so the claims stay precise.

PhD direction

The later notes point toward hierarchical attribution, intervention validation, and LLM factuality/style attribution.

Start by defining what attribution is allowed to claim.

A strong attribution project begins with clear units, clear utilities, and careful claims.

Read the first note
中文版本