OPUS selects training data in optimizer space and builds a 30M-token benchmark proxy

VoidAsuka · x · 2026-07-25

The authors describe OPUS, a data-selection method that scores examples in the geometry actually used by the optimizer instead of raw gradient space.

They also introduce BENCH-PROXY: rather than using benchmark validation data directly, they embed the benchmark, retrieve similar documents from the pretraining corpus, and build a 30M-token proxy pool to keep the target direction on the pretraining manifold and reduce noise.

Original post →

More from Research

Research channel →