Open Models Almanac
Methodspost-trainingsupervised fine-tuning

specific method · filed under post-training

Rejection sampling

Rejection sampling is a data-filtering method that turns reward functions into filtering criteria to build a high-quality training subset.1 It keeps the highest-quality sample from multiple independently generated candidates, reweighting the training distribution toward better instances.2

Trade-off It improves sample efficiency and training stability, but some filtered-out trajectories still contain valuable high-quality content.3

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. the reward functions used in RL are transferred into a data filtering pipeline to construct a high-quality training subset — GLM-5: from Vibe Coding to Agentic Engineering
  2. We adopt a Best-of- selection strategy, in which the highest-quality sample is retained from multiple independently generated candidates — GLM-5: from Vibe Coding to Agentic Engineering
  3. leading to improved sample efficiency and enhanced training stability — GLM-5: from Vibe Coding to Agentic Engineering

Also called rejection sampling fine-tuning.

sources
2
models
2
labs adopt it
2
strongest
used
↑