4,006 samples

The '4,006 samples' dataset is part of a larger collection used to evaluate the safety and behavioral equivalence of speculative decoding outputs against target model outputs in the context of temperature-zero inference.