Open Models Almanac
Methodspost-trainingrollout & RL infrastructure

single mention · post-training

Discarding early-returned short samples

Discards early short samples to smooth the transition to the steady-state length distribution and reduce overfitting to short sequences.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See rollout & RL infrastructure for the methods filed beside it.

Filed under post-training › rollout & RL infrastructure