SLAP: Stratified Loss-based Pruning for On-Policy Data-Efficient Instruction Tuning

In three linesSLAP is a batch-aware data selection framework for instruction tuning that evaluates learnability at batch composition level rather than individual samples. Using stratified sampling and relative distance optimization with Hessian-approximated gradients, it matches full dataset performance with 20-40% less training data across LLaMA, ChatGLM, and diverse tasks (dialogue, translation, QA).

Summary generated by Claude — human-verified