ConceptRejection Sampling Fine-Tuning / RFT / 拒绝采样
Rejection Sampling Fine-Tuning
A post-training paradigm that does not require human preference annotations. The model samples multiple outputs for the same input, filters incorrect samples using external criteria, and applies supervised fine-tuning on the remaining successful trajectories. Compared to RLHF, this approach incurs lower implementation cost.
Timeline (last 90 days)
Oct 24
JOLT is not the first work to attempt to compress post-training costs; previous directions include rejection sampling fine-tuning, the direct preference optimization (DPO) family of methods, and various self-rewarding mechanisms.
Unverified50%