Group-Level Training Data Attribution with Exact Shapley Analysis

This project asks a data-centric question: which interpretable groups of training examples are responsible for different aspects of model behavior after fine-tuning?

The method defines auditable training-data groups based on written-standard labels, enumerates all feasible data coalitions, trains models across the coalition space, and computes exact group-level Shapley values for multiple utility functions. The evaluation separates translation quality, terminology accuracy, and written-standard behavior instead of reducing data contribution to one scalar score.

My contribution was the research framing, coalition protocol, Shapley computation design, model training setup, metric design, random size-matched baselines, and cross-architecture validation plan. Because the work is still manuscript-stage, the site presents the method and contribution at a high level.

• Formulated group-level data attribution as an exact Shapley analysis over interpretable data groups.
• Enumerated all 16 coalitions for four written-standard groups and evaluated multiple behavior-specific utility functions.
• Compared true groups with repeated random size-matched groups to test whether attribution patterns are reducible to group size.
• Extended the protocol from encoder-decoder MT to decoder-only instruction-format LoRA validation.
• Used bootstrap confidence intervals and manual audit planning to make the attribution analysis more robust.

Size-Matched Baseline Diagnostic

The project compares true linguistically defined groups with random size-matched baselines to separate group identity effects from group-size effects.