https://wezteoh.github.io/tags/llm/ Recent content in LLM on Some days I delve https://wezteoh.github.io/%3Clink%20or%20path%20of%20image%20for%20opengraph,%20twitter-cards%3E https://wezteoh.github.io/%3Clink%20or%20path%20of%20image%20for%20opengraph,%20twitter-cards%3E Hugo -- 0.148.0 en-us Sun, 12 Jul 2026 00:00:00 +0000 https://wezteoh.github.io/posts/understanding-sft-dynamics-around-the-best-pre-rl-checkpoint/ Sun, 12 Jul 2026 00:00:00 +0000 https://wezteoh.github.io/posts/understanding-sft-dynamics-around-the-best-pre-rl-checkpoint/ <blockquote> <p>A recent study showed that when training reasoning models, the best SFT checkpoint for downstream RL can occur well before the checkpoint with the strongest SFT performance, with validation loss emerging as a surprisingly effective predictor of post-RL performance. In this post, I reproduce this observation on a much smaller language model and examine how the model evolves around the optimal pre-RL checkpoint. Along the way, I explore several training dynamics that may help build intuition for why validation loss within a training run has the capacity to predict downstream RL performance. Github link coming soon.</p>