SDO framework accelerates LLM post-training with adaptive data batching

Post-training large language models is notoriously expensive, and current efficiency tricks focus on picking informative samples or tuning schedules—not on how data is organized. SDO challenges that static preprocessing mindset with an exposure-driven feedback mechanism. Instead of fixing data partitions before training, SDO rearranges mini-batches epoch by epoch using frozen external embeddings, so no model warm-up is needed. Within each epoch, locality-aware batching groups similar samples via KNN neighborhood traversal, creating coherent mini-batches. Across epochs, exposure-balanced scheduling tracks how often each sample appears and lowers the sampling probability of over-exposed examples, preserving long-term coverage without ever dropping data permanently.

SDO is designed as a drop-in data organization layer for any post-training pipeline, and tests show it accelerates convergence across SFT, DPO, and GRPO—with the biggest wins in the early-to-mid training phase. The approach produces more coherent gradients and more balanced accuracy across different question types, directly attacking the common problem of redundant updates for some samples while others remain under-optimized. Because it operates on embeddings and doesn't require model changes, SDO could be a cheap, practical upgrade for teams fine-tuning LLMs in production.