DRIFT Refines LLM Instruction Data for Peak Performance

Optimizing the training data distribution is paramount for maximizing the capabilities of Large Language Models (LLMs through Supervised Fine-Tuning (SFT). While many data curation techniques focus on accelerating training under budget constraints, the challenge remains to refine data to push the model's ultimate performance ceiling. DRIFT addresses this by employing instance-level data attribution using Influence Functions (IF). It identifies and overcomes two key limitations of standard IF formulations: a "proximity gap" arising from off-policy validation targets and a bias towards gradient norms. The framework utilizes the model's own on-policy rollouts as validation targets, which minimizes the parameter proximity gap and aligns better with IF's local neighborhood assumption. It further applies signed weighting based on trajectory correctness and debiases influence scores. This allows a small set of validation queries to reliably attribute the full dataset, consistently raising the performance ceiling for 7B-parameter instruction and reasoning models, outperforming existing data curation baselines.