Report
SFT can match RL’s generalization if its data budget catches up, @gurtej__gill_ says
Summarizing Qwen experiments, @gurtej__gill_ says larger, more diverse datasets let supervised fine-tuning rival reinforcement learning.
TLDR
In Qwen2.5 and Qwen3 experiments, supervised fine-tuning matched reinforcement learning’s generalization when its training-data budget matched RL’s rollout budget—and beat it when data sources were also mixed, according to @gurtej__gill_’s paper summary. The catch, @srchvrs argues, is that diverse data isn’t easy to collect; generating it may be part of RL’s advantage.
Combined views
129
3 Sources, first seen ago
Comments